I created a survey to measure student perspectives on different subjects in Higher Education. In the survey, the students are asked what subjects they have experience with, and they will only be asked questions on the subjects that they selected. For all subjects, there is an identical series of likert scale questions asking about learning, how welcoming it is, stress, and more. I also ask about the gender and race of the participants to see if there are any differences depending on the demographics.

Now, the issue is that most participants only have experience with some of the subjects and not all of the options, and therefore there is a lot of “missing data”. For example, participant A has taken sociology and mathematics and answered questions for both subjects, but their responses for the rest of the subjects are empty. Meanwhile, participant B has taken computer science and economics, so they could not respond for the rest of the subjects.

It results in a table like this (abbreviated version):

Psych_welcome Psych_learn Psych_stress Socio_welcome Socio_learn Socio_stress Math_welcome Math_learn Math_stress CS_welcome CS_learn CS_stress Econ_welcome Econ_learn Econ_stress Gender Race\Ethnicity
Participant A 4 2 5 3 3 4 M Asian
Participant B 3 5 5 4 3 4 F White

Initially, I thought about using repeated measures MANCOVA to see if there are differences between class subjects and possible relationships between measures like learning and stress, along with covariates such as gender and race, but since there is a lot of missing data, I am not sure if that is possible.

I also thought about editing the table so that the class subjects became one of the columns, but then each participant would take multiple rows and we would not be taking into consideration the fact that many rows would have been answered by the same participant.

Thanks for the help, and please let me know if I didn’t explain anything well enough!


1 Answer 1


Welcome to the site.

This is not an easy problem and you probably don't have enough data to answer it well. It is equivalent to a problem known as "test equating" in psychometrics and it needs a LOT of data to do really well.

Still, you could look that term up and see if there is something that you can do, at least a little. I studied it in grad school, but that was 30 years ago and I've forgotten a lot (and there may be new methods).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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