Hi,
I have a question regarding SPSS and the backwards step selection method. I have several variables but want to keep four of them constant. Does anybody know if you can fixate four variables to remain in the backwards step selection regardless of the p-value?
Many thanks for considering my request.
Kind regards Stine
You should not be using stepwise methods at all. They are notoriously unreliable. See https://www.stata.com/support/faqs/statistics/stepwise-regression-problems/
You can use the non linear regression in SPSS as you will write the model you want.
Hello Stine,
I agree with Ronán Michael Conroy that there are many reasons to avoid any step-based model (not the least of which is that there is no guarantee that you will end up with the best ensemble of IVs to account for differences in the DV).
However, what you propose is a bit like a hierarchical model, in which you decide (usually, for reasons based on theory or prior results) that some variables should be forced first into the model, then the potential efficacy of others to (potentially) improve the explanatory power is evaluated. If you absolutely, positively must pursue the backwards elimination process via spss, then you could: (a) residualize the DV for the four chosen IVs; and (b) use the model residuals as the DV for considering the subsequent IVs as candidates. That would accomplish your query goal.
But, step methods are notoriously bad, and are opportunistic, so you'd definitely want to validate your resultant model with new/different data.
Good luck with your work.
Ronán Michael Conroy, Abdulmuhsin S. Shihab, David Morse Thank you so much for your answers! I appreciate them all.
As you all have touched upon, variable selection of this kind is often unwanted. However, my problem is that I am working with a minimal sample size (n=27) but got a large number of variables (27). The utilization of step methods was an attempt to (somewhat) accommodate the rule of variables per event. However, I hear that you all strongly recommend me to go another way.
David Morse, you mention a hierarchical model as a better option. Would the best approach be to add the variables in blocks according to their theme? For example "mental health", "physical health", "sociodemographic", "psychosocial work environment" and should the four variables I wanted to fixate/force into the model be the first or last to be added into the hierarchical model?
I appreciate any advice you can give me. Thank you in advance.
Kind Regards Stine
Hi Stine, there are a couple of points that virtually everyone has tried to make that you seem to have missed. The most important is that if you run stepwise methods repeatedly the selection doesn't repeat. U get almost all possible selections. How can you possibly know which is best.? See the discussion in the first two sections of the attached. Now I was trying to help a PhD student with a similar situation a few months ago. In her case n was about the same as p. After trying everything I could think of including the methods in attached paper that if n and p are about the same there's no method that will help you because you don't have enough information to pin down an appropriate predicion equation. An easy way to convince yourself is to try a few examples with n and p approximately the same. The idea is that one point per regression coefficient is not enough information to pin down the coefficient. Thus you're unlikely to be successful with the sample you have. More data or a more parsimonious model would help. BTW we tried p g.t. n models too from microarray papers these didn't work either . In this case you need MORE information than you have from somewhere. Statistical methods require enough information before they will work..mean of a sample of 1 is equal to the point that you have. Good luck, David Booth
In a non linear regression in SPSS you can write a linear model and put the constant you want and made many trials enter all the variables you want and then tries another models by cancelling one Variable at each step.
I add that you can fix the four variables you mention with a Coefficient equal to one.
Whoa – with 27 variables and 27 participants you have actually one more participant than you need to achieve an R2 of 100% an adjusted R2 of zero!
Your problem is that any model using random variables will end up predicting a dataset so long as it has the same number of predictors as the number of cases. The intercept is a predictor, so only N–1 random variables are needed for perfect prediction.
The adjusted R2 of zero is because when you adjust R2, you adjust for the proportion of variance that would have been explained by a model of random predictors with the same number of predictors as your model.
So there is no solution here. You just don't have enough data. If you hypothesis was so complicated that it required 27 variables to specify it, you needed a much larger dataset to test it. I would start again by drawing a diagram of your hypothesis to see what the key variables really are.
But bear in mind that with only 27 observations your power to detect associations is very low indeed, and that stepwise methods will simply fit random variation in your data.
This is the danger of ResearchGate. There are people out there telling you how to do something better that is actually the wrong thing to do.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS