The questionnaire I worked on consists of 49 items rating on a 6 point likert scale. I found that most participants chose the highest point of many items. However, the results of CFA are good? Is it named ceiling effect? what are the effect of it on the analyses?
When data are highly skewed, this can "attenuate" the observed correlation so that is not possible to reach the usual maximum value of 1.0.
Personally, I would say that if you get a good fit with CFA, that is all that most people would care about -- when is the last time you saw someone publish a table that included not just the means and std. deviations for their variables but also the skew and kurtosis?
Also, some Structural Equation Modeling programs such as MPlus allow you to make alternative assumptions about skewness, but I am not familiar with the technical details that involves.
When data are highly skewed, this can "attenuate" the observed correlation so that is not possible to reach the usual maximum value of 1.0.
Personally, I would say that if you get a good fit with CFA, that is all that most people would care about -- when is the last time you saw someone publish a table that included not just the means and std. deviations for their variables but also the skew and kurtosis?
Also, some Structural Equation Modeling programs such as MPlus allow you to make alternative assumptions about skewness, but I am not familiar with the technical details that involves.
Forough,
There are at least two aspects of what is going on here.
Yes, these are clearly ceiling effects. But where you go from here has more to do with modeling these items as ordinal vs. continuous in your CFA.
A ceiling effect can reflect, for example, a censored-normal distribution. This is a continuous-data model where it is assumed that many of the 6s would be higher if the scale went that high.
In ordinal-data terms, it simply means that the thresholds in the cumulative probit or logit model are relatively low.
In this situation, I'd recommend a categorical-data CFA. If, on a given item, most of the respondents are at 6, I doubt you can reasonably argue that the scores below 6 are continuous and interval.
Using interval-data methods for a 6-point categorical-response scale can be argued, but most advocates for doing so include a qualifier that the distribution should be at least vaguely symmetric.
Pat
Thank you Pat
For curiosity, I had the data checked by a statistician .
hi
these publications may help:
http://kueprintskh.kingston.ac.uk/17226/1/Bowling-A-17226.pdf
http://dare.ubvu.vu.nl/bitstream/handle/1871/32288/271217.PDF?sequence=2
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Applied Statistics
- Survey Design
- Questionnaire