0
$\begingroup$

I would like to rank independent variables in a logistic regression model based on relative importance. I've read about standardizing the variables prior to entering them in the model. So in this context, how can I standardize a categorical variable with 5 levels.

My final model has mix of continuous and categorical variables. If standardization is not the right approach for this problem please suggest me alternate approach.

$\endgroup$
  • $\begingroup$ What is your purpose in trying to rank-order the independent variables? This is often not a useful exercise, as the regression coefficient for each of the independent variables in the logistic regression is effectively adjusted for the influences of the other independent variables in the model. So an "important" variable in your full model might seem unimportant in a model with fewer independent variables, or vice-versa. $\endgroup$ – EdM May 14 '15 at 19:09
-1
$\begingroup$

It is not possible to standardize a categorical variable since there is mean nor variance defined on a categorical variable, even if it is ordinal. Instead, I would perform the model fit with the continuous and categorical varialbes within and check the statistical significance of each single predictors using TYPE III tests (see R command anova(fittedModel, test="Chisq") as example ). TYPE III tests assess the predictors' imporance by verifing if the increase in deviance residuals of the model fitted without the tested predictor is statistically significant.

$\endgroup$
  • 1
    $\begingroup$ Type III tests as demonstrated by Venables and Therneau have serious problems with them. I suggest using ordinary likelihood ratio $\chi^2$ tests ("type II tests") or ranking variables by partial Wald $\chi^2$. My course notes have an example where I use the bootstrap to get confidence intervals for ranks of predictors using Wald statistics corrected for d.f. See link to handouts from biostat.mc.vanderbilt.edu $\endgroup$ – Frank Harrell Jul 26 '15 at 12:53

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.