I don't think that there is a threshold. You do need to consider "why" you have so many missing variables in those two variables. If there is a reason that does not go into the MAR or MCAR categories, then you may be better to explicitly model "missingness" with an indicator variable.
Personally, I would be concerned that either of them fit the PROC MI assumptions. Skipping can be because the person didn't see it or something else extraneous (MCAR), but can also be because the person found the question intrusive (income, for instance would not be appropriate for PROC MI). Not Applicable is definitely an answer that needs to be modeled, using PROC MI to impute another value is going to bias the results.