Is there a way to analyse missing values of nominal data. I heard about multiple imputation, but for my understanding, this provides temporary values to the missing value. But all I want is to identify patterns of missing values, and understand if there is really a pattern on the said missing values (that may imply something about the way the questionnaire is presented or the structure of it) or the values are just missing randomly.

Sorry if my question is a bit confusing. This type of analysis is all new to me and would be grateful for all your help!

  • $\begingroup$ The first thing to do is to look at frequencies of missing values per question and per participant (you may have done that already). This will show you whether specifc questions and participants are particularly affected. Further steps depend on what kind of pattern you are particularly interested in. For example you can relate numbers of missing values per participant to variables such as age and/or other (non-missing) variables of interest. You can also look whether missing values on particular questions can be explained by other variables. $\endgroup$ – Lewian Jan 11 '20 at 20:07
  • $\begingroup$ Dear Lewian, thank you very much for your comment/advise. I wonder if you could provide further advise as to what sort of analysis shall i do to be able to figure out if missing values per question and per participant are related? Sorry this is all new to me. I would be very much grateful for any advise you could provide $\endgroup$ – Shin04bix Jan 14 '20 at 13:35
  • $\begingroup$ One thing that could be helpful (assuming that there's a manageable number of questions and participants) would be a plot in which you have questions on the x-axis, participants on the y-axis, and points or red marks or whatever if a question is missing for a participant. Probably better with hierarchical clustering of participants and questions as done for standard heatmaps in R (unfortunately I don't have the time to tell you in detail how to do these and it may require some tedious figuring out). $\endgroup$ – Lewian Jan 15 '20 at 0:47

You are right that multiple imputation assumes that the data is missing at random and will not help you uncover these patterns.

One way to see if there is a pattern is results oriented - because you have nominal data, you can add a label to represent "missing data" and feed that into your learning algorithm. Then you can also use an imputation method to deal with missing data, feed that into the algorithm, and see which approach gets better results. If the first one does better, there might be a pattern.

Other approaches would include data exploration - can you characterize who left out what value? But I don't know your data, so I'm not sure I can help there.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.