Hi, I have an Independent Variable "Perceived Discrimination" measured on Likert Scale (5 questions), and Dependent Variable (Customer Complain Behavior). The DV is not measured on Likert Scale, rather respondents are given a choice to rank 6 options (in order of their preference).
1. Seek Correction, 2.Ask for explanation, 3.Seek Apology, 4. Express Anger, 5. Seek Compensation, 6.Seek Redress
Some respondents recorded their response as (1,6,3,2,4,5) others (5,4,3,1,2,6) and so one ... every respondent has given his own order of preference.
How can I related the "Perceived Discrimination" which is Likert Type data with "Complain Behavior" which is ranked data?
Hello Farhan,
What you might try is this: Compute Spearman correlation between the IV (summated or averaged across the five questions) and the rankings given to each of the six response options, separately. Then, you can discuss whether the resulting correlations appear stronger for some response options than for others. Here is a web site that helps with statistical comparisons of dependent correlations: https://www.psychometrica.de/correlation.html#dependent
A problem with trying to evaluate values for all six response options simultaneously is that there will be linear dependence among the values (e.g., If I know the rankings someone gave to options 1-5, I can deduce the 6th ranking). As well, the correlations among the options will tend to be negative (e.g., if I rate one option high, the others by necessity are given lower ranks). These are issues with forced-choice approaches, or ipsative, measures in general.
Good luck with your work.
I think I would probably consider modeling this as a multilevel ordered logistic regression with the six options nested within person and the scale and option as predictor. This would also allow you to model the interaction between scale score and option.
Regression procedures for ordinal variables.when i was in grad school, we used LISREL. I have not kept up with these procedures.
These days I use R so I'd probably use the ordinal package with the clmm() function or brms (a Bayesian package that can fit a wide variety of models in eluding ordinal ones).
Explaining a set of choices on a Likert scale by a multinomial or an ordered logistic regression would be a good choice, but this is impaired by the structure of your IV which is an unordered set of 6!=720 possible responses.
I would recommend to first (try to) reduce the "dimensionality" of your IV.
Perhaps you might first try to describe the "differences" between subjects by any index of similarity/dissimilarity.
examples :
- a binary distance : are the ranks of the answers all equal or not?
- a semi-quantitative distance : the sum of absolute differences of ranks between subjects,
- or ... (a look in the literature on dissimilarity indices ?)
And then try to accommodate a bi-plot to describe the matrix of distances between subjects. Then you will have a set of quantitative variables describing" the subjects replacing you in a position more favorable for logistic regression.
But the question of the interpretation of the "new" subject coordinates is left open.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis