Logistic regression - calculating feature importance

#1
Hi there,

I am running a logistic regression and struggling to understand how to state a relative importance of each of the attributes within the model.

I have the 'estimates' (I think this would be the 'b' values in SPSS), but my variables have different scales (some are binary and some are 7 point scales).

How would I move from the 'estimates' to something like a '% contribution' to the model? One thought I had was to use the odds-ratios (and looking at their relative contribution), but do I also need to consider the different scales in this (e.g., by multiplying the odds-ratios but the range of the respective variables)? Or this approach not correct at all?

Any advice would be really appreciated!

Many thanks!
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
What would be the purpose of this?

When a person moves away from linear regression with standardized covariates - things get hard and blurry. Say you had two covariates:

X1 OR: 1.5 (95% CI: 1.1, 1.9)
X2 OR: 1.3 (95% CI: 1.2, 1.4)

Which is more important? Well X1 has a larger effect estimate and X2 is way more precise. If doing Bayesian analyses you could see where the middle density is, but both covariates are in the model at the same time. Is there variance inflation and if so why. Also, the estimates are conditional, and the interpretations get tricky when thinking about the base cases. Also, without holdout data do you even know if these results will generalize?

I would just recommend plotting the ORs with confidence intervals and let the reader decide their interpretation. Remember to use a log based axis, so they are scaled correctly since an OR of 5 and 0.2 have the same magnitude of effect on the DV.