I have a questionnaire that contains 5 likert scale as a response for all items. some items are (never-always) and others (strongly disagree- strongly agree).
during manipulating data, i found that the composite score for the variable that measure dependent variable is not distributed normally, even after transformation using square root, log10 or inverse.
can i collapse 5 likert scale (never-always) in order to have normal distributed data?
**(note: actually this scale were adopted from previous questionnaire with yes/no response but i modified it to 5 likert scale to avoid confusion for respondents, in addition to increase the options of way of analysis)
There seem to be (at least) two issues here:
Heiko Breitsohl thanks for answer
the original scale from where I adopted these questions were yes/no
so i thought that if i converted the scale from 5 likert to yes/ no , it may give normal distribution of data
or may I convert from (never- seldom- sometime- often - always) to 3 group for example (never- seldom & sometime - Often & always). I see thesis for master student do this (the last categorizing above), but i'm not sure is it right or no !
I see two issues with the attempt to reduce the number of categories:
- By definition, the Normal distribution is continuous, so reducing the number of categories will never get you closer to normality. In other words: the smaller the number of categories, the farther away from normality, particularly with only five categories to begin with.
- Reducing the number of categories will almost inevitably lead to a loss of information because you are unable to distinguish between cases that used to be in different categories but are now in the same one. I agree that a lot of people do this, but usually without sound reasons.
So, reducing the number of categories will probably not help you. I would recommend trying to handle non-normality in a different way. Depending on the model you want to use, you could use a robust estimator or you could switch to a model that explicitly treats the data as ordinal (such as ordered logit regression or ordered probit regression).
No. There is no sufficient evidence that PLS would help with that.
See
Article On the Adoption of Partial Least Squares in Psychological Re...
I am afraid that there is just no magical solution.
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India