Most of the justification for the use of mean on ordinal data is common practice in social and psychological studies. Are there any stronger justifications for use of means? i want to use the mean values to further do a regression analysis, would that add further error?
Dear Abel,
For regression analysis, it is in the first instance important to identify WHAT you are doing and WHY. so without specific knowledge of your data and purpose of the analysis it's difficult to advise.
However, in general, using mean for ordinal data is indeed broadly practiced; however personally I am not really a great fan of it. I prefer to check the relationships between the variables using dummy-coded variables for each ordinal level or high/low scores. It provides a better idea about the nature of association, and whether indeed the use of mean is justified.
Would that help?
kind regards, Marija
Hi Abel,
W should admit that we ALL do this. Those of us who went through a typical schooling system anywhere in the world, had their GPA calculated. GPA is ... an average, and it is based on a very questionable assumption of equal intervals between grades - assumption that difference in knowledge learned/skills acquired between "A" grade and "B" grade is the same as between "C" and "D".
Trying not being a major hypocrite ;-) (having calculated students' GPAs myself many times), I would see this problem in more pragmatic perspective: is the model you built useful? Can you show that it predicts whatever you are trying to predict not only in theory (since breaking some assumptions may have invalidated the theoretical grounds for your predictions)? Maybe try using the approach offered by machine learning crowd - dividing your data into train and test, if you build your model on "train" data and then you can successfully predict the "test" subset using that model, it may give you a leverage in showing that even though you broke the theoretical assumptions, the model still works.
Two more perspectives that may be of some use:
- remember that binary measures do not break the assumptions of equal intervals, so you could maybe collapse your ordinal data into binary sets?
- if the data refers to some psychological traits, there are some publications which show that humans tend to "assume" intervals are equal if presented with more than 4-5 categories to choose from (which kind of relaxes the problem of the GPA calculation, described above).
- Badges
- Science topic
- Similar topics
- Social Science