Handling survey option "I don't know"

I am at the early stage of analyzing a survey. Most items have ordinal
scales (Extremely useful, very useful, somewhat useful, etc.) but some
also offer the "I don't know" option.

When reporting relative frequencies, I can leave the "I don't know"
category in, since that information is interesting for those who will
be reading the report. However, I would also like to use polychotomous
logistic regression to take advantage of the ordinal nature of the
variable. Is this something I can do and if so, how do I handle the "I
don't know" option?

Finally, would anyone recommend one or more books that would answer
such practical questions?

Thank you,

Daniel

.