Relative contribution of variables in logistic regression

#1
Hi, I have been using logistic regression to determine what influences the chances of a person becoming unemployed. A number of variables have all proved highly significant (P<.001) but I would like to determine the relative contribution made by each variable so that I can say which is most important.

I have read that relative weights analysis can b used to determine the importance of independent variables but this is new to me and I don't know how it applies to logistic regression or how it is implemented in software packages - in particular, SPSS.
 

jpkelley

TS Contributor
#2
Standardized regression coefficients...different ways to do this, depending on what kind of variables you have (continuous, etc.). This simply involves z-transforming (mean=0, variance=1) each variable and then running the regression. All of your effects are then in standard deviation units. Each effect size (beta parameter) can then be compared to one another. A couple of Google results:

http://en.wikipedia.org/wiki/Standardized_coefficient
http://dss.princeton.edu/online_help/analysis/regression_intro.htm

Hope this gets you on the right track...
 

trinker

ggplot2orBust
#3
In logisitc regression the odds ratio is usually the measure given as the strength of effect of the IVs (including group membership). In SPSS I think this is in the output as exp(b) or something similar.
 

spunky

Smelly poop man with doo doo pants.
#4
Azen & Traxel (2009) extended a technique called "Dominance Analysis" developed by Azen & Budescu in the context of OLS multiple regression to determine relative variable importance. i believe it can be done now in logistic regression and i'm pretty sure Azen programmed an SPSS syntax macro somewhere out there for people to use this technique to determine relative importance of predictors. just google "dominance analysis" and "logistic regression" and it's the first or second search result that you'll get...
 

trinker

ggplot2orBust
#5
Spunky that was informative. Thanks for the post. My suggestion is the stardard convention for strength of effect but doesn't really tell you about the relative improtance of each variable. I haven't checked out the article yet but as soon as I post this I'm going to go get it.