"Does not apply," and "missing" are two different things.  (Also, "zero" is a third 'thing.') 

[If something is "missing at random," it is an "ignorable nonresponse," meaning there is no hidden mechanism, so that, say, for continuous data, a mean will work, though it should not be used as real data when calculating standard errors.  If you have a "nonignorable nonresponse," you need more information.]  

But "does not apply," if that is what you intended, appears to mean you are outside of the population of interest, and it should in no way be considered. 

PS - I don't know how your software is laid out, but if you have, say, 811 respondents,  and 12 are "NA," then you really have 799 respondents. 

"Does not apply," and "missing" are two different things.  (Also, "zero" is a third 'thing.') 

[If something is "missing at random," it is an "ignorable nonresponse," meaning there is no hidden mechanism, so that, say, for continuous data, a mean will work, though it should not be used as real data when calculating standard errors.  If you have a "nonignorable nonresponse," you need more information.]  

But "does not apply," if that is what you intended, appears to mean you are outside of the population of interest, and it should in no way be considered. 

PS - I don't know how your software is laid out, but if you have, say, 811 respondents,  and 12 are "NA," then you really have 799 respondents. 

Are you looking at each item individually or are you wanting to form a scale?  If the former, just report the percent who answered NA.  If the latter then use an average score by using the number of items without an NA response.  You will need to determine how many items need to have been answered in order to create a scale score for that respondent.

How many not applicable and how many actual missing? As James nicely says, there are different types of ``missing'' and these will arise in different ways, and these different ways correspond to different ways to model them. One thing I was wondering, if an item is not applicable to some one, is the entire trait not applicable (say you have 5 traits/dimensions, and one isn't applicable to left handers). This seems different than if there are just a couple of items in each trait that are phrased in such a way not to be applicable to some.

Thank you for your answers, the item with most DNA answers have around 15 that signed the option. The measure is about quality of working life. So, some of.the items actually do not represent an aspect of quality for every sector. One option that i can take maybe is to exclude items with excessive DNA justifying according to content. The way it is right now the sample drops by a huge amount due to DNA answers. 

Yes you should have a cutoff for missing data even if it is because it is not applicable.  Decide what the minimum number of items you need to have to analyze and drop respondents who drop below this threshold.  

You can provide data by sector.  

Not applicable does not mean missing. In a survey I am looking at, once someone says "no" to a certain category of questions, they no longer are asked the proceeding detailed questions in that category. For example if someone chooses their job status as  "they are not working", then the question proceeding it for "number of days they have taken off  work" is not even asked and a "not applicable" is automatically filled in their cell. Essentially this would mean that they were never included in the cohort and were not treated the same as the rest of the people who said yes. You would have to drop them from your final cohort if you are analyzing the response to that question.  

I would advise you to review all questions that you are interested in to make sure that they never had "not applicable" as an option. Once you have a "not applicable" as an option you would have to see how the survey was designed to see if the question was asked to those people or not.

Hope this helps :)

As you reffered to Winsteps, I presume you want to do a full item analysis of a set of items making up a scale (Rasch model?). You might want to try instead RUMM2030 software, as they have the option to analyse item data with missing responses for some persons. A strategy might be to analyses complete data, then data with missing (maybe other packages does as well?). If there is no great difference, in your results, it should be fine to use the analysis with the missing data - of course, it all depends on have many missing you have.

As I understand your description the NA option should be trated as missing, as the persons using this cannot answer the item(s) in question, and as such this data should/could not be included.

hope this is usefull

Tine   

As mentioned above, Omit is definitely different than Not Applicable or Not Administered.  You need software that can handle such data.  I totally agree that dropping respondents is not really an option, nor is using CTT.  RUMM might be an option.  Try Xcalibre too.