relative impact of predictors


Fortran must die
I am frequently asked to rank the relative impact of predictors on a dependent variable based on the results of a regression. That is, does this variable influence (cause change in) the dependent variable more than another predictor.

For linear regression I judge relative impact by standardized coefficients (beta weights) with a larger number showing greater impact. For logistic regression I use Wald Chi Square values (there appears to be no agreed on standardized coefficient for logistic regression, I think I found 14 different ones last time I looked).

Even for beta weights there is a problem some feel with ranking impact this way if you have dummy variables (as I do).

Any comments on this would be deeply appreciated.