Rather sounds like you want to treat it as missing some times and not others. That calls for multiple variables.
There are a several approaches to attack the problem. 1) and 2) model the effect of missing explicitly and are reasonable when there is moderate missing data so the model has enough power. It is particularly valuable if the missingness is not at random. 3) applies different principles and is much more useful when the presence of missing is more of a nuisance.
1) Leave it as 3 levels and treat it as a class variable.
2) Equivalently, and I find this to be more interpretable.
-- create a new variable IncomeMiss that is coded 0/1, with the 1 as missing and 0 otherwise.
-- recode income to 0/1 with the level of interest being the 1, and 0 otherwise.
-- put both variables in the model as continuous.
Then the effect of Income level 1 is measured by the coefficient for income and the effect of missing income is captured in the other variable.
3) Apply PROC MI and MIANALYZE.
Doc Muhlbaier
Duke