I have conducted a human intervention study and measured various physical and metabolic characteristics pre- and post-intervention. I would like to conduct stepwise backward regression in SPSS to determine which variables best predicts the change in another variable in a bid to explain my results further.
I have measurements pre and post and I have calculated the actual change in the variables. Is it correct to perform stepwise regression using the calculated change in each variable or how do I add this type of data to SPSS for correct analysis?
As a random example, do changes in body weight and body fat explain changes in TNF-a concentration?
Dear Kiera,
You need to decide which variables should depend upon others before you can select an appropriate model. For a multiple linear regression, there should be one scale dependent variable (DV) and multiple independent variables (IVs) which can be either scale or binary.
From your description, it is hard to suggest what your DV might be. Your IVs could be the difference between your pre and post intervention scores, or they could be both pre and post scores as separate variables.
You have mentioned using stepwise backwards regression. This is a good choice, but the final model should still be validated. There are also assumptions that need to be checked.
To expand on Peter's suggestion, based on your example TNF-a concentration is the dependent variable. I am still trying to understand why you want to take the mean of the changes. The difference of values in pre and post intervention (for variables like body weight, body fat etc.) for each subject would be your independent variables. Then you check for the variable importance through the p-values in stepwise regression.
Thank you all very much for your help and recommendations. I really appreciate it!!
My apologies I didn’t mean to say the mean, it’s the difference between the pre and the post measurmernt for that variable.
I’m wondering whether I calculate the pre to post change in each of the measured variables (DV and IV) and add them to the SPSS data sheet. Then use the column of change in TNFa as the dependent variable and the other columns (e.g change in body weight and body fat) as the independent variables and run it as normal? Thank you very much!
If your independent variables are not independent (correlations different from 0), the least squares regression is not always pertinent. For example, you may have different results performing a backward or forward selection.
In fact, the result is function of the order in which you introduce or remove independent variables from your model.
If the correlations between your independent variables are too high, you could perform a regression on principal components are, better, a PLS regression.
I don't use SPSS, but I'm sure there exists PCA and PLS in this software.
To answer your initial question, there is not a correct value. It's to you to decide, and, of course, precise in your report. Very usually, it's 5% or 10%.
Hope this can help you,
Pierre Chagnon
Stepwise regression is akin to a 'fishing expedition' and should be avoided if possible - you should use theory and exisitng research to guide variable selection
Just a piece of advice for after you are sure of what are IV and DVs. Avoid stepwise regression. It is very unstable. See the attached for an introduction to the topic and what you can use to replace stepwise. Best wishes, David PS only use variable selection if you are building a predictive model.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS