Background to problem I am currently in the process of computing some quantitative data (Questionnaire likert scales) and there is clear differences in missing data on a specific item ~400 missing responses, compared to ~100 (on the other 9 items).
Participants were provided with comment boxes, so they can add qualitative information if they wished.... 70% of respondents who did not answer the quantitative component did provide comments. A content analysis of comments for that item revealed that around 70% of those who commented (and about 50% of the total missing data) said that the question was not applicable to them. While another 20% (15% of total missing data) answered the question with odd and slightly irrelevant information - suggesting to me that they didn't understand the item and took punt at providing me an approximation of what they thought was related information...
Problem What do I do with MNAR modelling when I know there are at least two reasons specific participants have not responded - Can I impute based on their commentary data - does this introduce subjectivity and potential bias?
Do I completely ignore the fact that I know why some people have missing data and run something similar this: http://journals.lww.com/epidem/Fulltext/2011/03000/Sensitivity_Analysis_When_Data_Are_Missing.25.asp and leave the software to fill in the gaps - does this lead to artificially inflated variance is non-responders? and does it work if there are multiple reasons for missing data?
Most modelling also looks at ratio data - I'm dealing with ordinal likert data...How applicable is sensitivity analysis is this situation?
Do i just ignore the question entirely - possibly introducing the biases associated with ignoring ugly data...
TL;DR What is theoretically the best way to deal with MNAR data when there are multiple reasons for missing data (~50% of missing data respondents felt the item was not applicable, ~20% appear to not understand the question, and 30% have not said why they didn't answer).
Thanks in advance