Hello Marvelous,

I agree with the recommendation of Vikas Ramachandra.  In logistic regression, the log likelihood statistic can be used for comparison of nested models. So, run the full model (all IVs), and note the -2LL value (in general, smaller is better for this value). Then, run successive models, each omitting one of the IVs.  The omit-one run which results in the largest increase in log likelihood indicates the (omitted) IV that contributed most (given the other IVs) to the model.  There are no guarantees, of course, that the same sequence of 'impact' would be observed in another sample.

Good luck with your work.

Hi, you can use the function calc.relimp from the library relaimpo.

https://cran.r-project.org/web/packages/relaimpo/relaimpo.pdf

Denis Cornet, thanks for the reply, however calc.relimp only works for linear models and not glms

Marvelous,

Have a look at "relimp" package.

It deals with class lm, glm, coxph, survreg, multinom, polr or gls.

you can turn off one covariate/predictor variable at a time (after the model is built), predict on a test set and see the error, then rank order predictor variables based on the errors.

If all your predictors are on the same scale (e.g., if they all have mean of 0 and SD of 1) then you can compare the standardized coefficients.

Hello Marvelous,

I agree with the recommendation of Vikas Ramachandra.  In logistic regression, the log likelihood statistic can be used for comparison of nested models. So, run the full model (all IVs), and note the -2LL value (in general, smaller is better for this value). Then, run successive models, each omitting one of the IVs.  The omit-one run which results in the largest increase in log likelihood indicates the (omitted) IV that contributed most (given the other IVs) to the model.  There are no guarantees, of course, that the same sequence of 'impact' would be observed in another sample.

Good luck with your work.