0
$\begingroup$

I am interesting in finding the relative importance of variables in a GLM model. The dependent variable is binary, while the independent variables are a mix of continuous and categorical. To do this, I would be looking at the coefficients with low p values from the GLM. The coefficients are standardized (and say uncorrelated).

My question is that do I need to assess model fitness before looking at the coefficients? I can, of course, look at predictive accuracy of the model, but the predictive accuracy might not be indicative that the individual coefficients are correct. I also thought of using R^2, but again it only measures overall model fitness. What are the recommended set of steps I need to take before I can trust the coefficients from a GLM are providing a correct estimate of relative variable importance?

$\endgroup$
  • $\begingroup$ To be more precise about the question, for assessing variable importance, is it enough to look at coefficients with low p-values or do we need to look at model fitness before looking at coefficients? $\endgroup$ – Rajesh Aug 31 '15 at 16:59
0
$\begingroup$

If I understand you correctly, you should assess how well the model fits before paying attention to the coefficients. If the model is a poor fit and the parameter coefficients are significant, then the coefficients are doing a good job at explaining a bad model.

Once you are satisfied the model fits well, you can interpret the coefficients.

If you are looking to figure out the relative importance of the predictors in your model, that's a different can of worms. The coefficient of each predictor is how much a one-unit change in it affects the outcome. But, critically, that effect is based a model that includes the other predictors you included.

This article explains some of the thorniness:

http://www.r-bloggers.com/the-relative-importance-of-predictors-let-the-games-begin/

$\endgroup$
  • $\begingroup$ Thanks for the response. The first part of your message addresses my question. To be more precise, for assessing variable importance, is it enough to look at coefficients with low p-values or do we need to look at model fitness before looking at coefficients? $\endgroup$ – Rajesh Aug 31 '15 at 16:58
  • $\begingroup$ I think you first need to look at model fitness. Then at the contribution of the coefficients. Low p-values aren't enough though, if you are looking for the relative importance. They just tell you the likelihood that you obtained a coefficient that extreme by chance. $\endgroup$ – bjsalami Sep 1 '15 at 1:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.