I am running logistic regression using glm in R on data, that has some indicator variables to it. Two of those have multiple levels and have been rewritten as (#levels-1) new dummy predictors respectively. As for variable selection, I use step (backward variable selection based on AIC), which is where I struggle:

Even though all dummies of the initial indicator were kept or removed, I wonder what one would do, if a single dummy (i.e. one of a set of associated dummies) was eliminated during the variable selection? After all, this would mean that I lose the information encoded for the base-level (since it is "indicated" by 0s in all the other dummies).

Could someone tell me what to do if such a case occurred?

Thanks & Regards

I am running logistic regression using glm in R on data, that has some indicator variables to it. Two of those have multiple levels and have been rewritten as (#levels-1) new dummy predictors respectively. As for variable selection, I use step (backward variable selection based on AIC), which is where I struggle:

Even though all dummies of the initial indicator were kept or removed, I wonder what one would do, if a single dummy (i.e. one of a set of associated dummies) was eliminated during the variable selection? After all, this would mean that I lose the information encoded for the base-level (since it is "indicated" by 0s in all the other dummies).

Could someone tell me what to do if such a case occurred?

Thanks & Regards