Hello,
- In our research, we used a scale that is valid and reliable. To check the factor structure, we used EFA and see that some of the items are not working (5 out of 24 items). Is it statistically acceptable to extract the items that are not working and use the scale? After extraction, thus with 19 items, we used CFA and the model fit is good enough..
- What should we do?
in your EFA, did you find one factor or several. And will you use factorscores in further analyses or do you want to use a sumscore? If you expect one factor you can use reliability analysis and test for additivity. It could also provide you with more information on the reasons why the five items do not fit. In my experience a one general factor (before rotation) shows up as a one-dimensional Mokken scale using a search procedure. But with Mokken I test whether a simple sumscore is sufficient (this is also done with a Rasch model where the sumscore is a sufficient statistic). Can you give a little more information that could make us understand what you have? Deleting 5 items is done often and most of the time it is allowed. But it also depends on the communalities. If they are very low, the items as designed do not measure as expected.
Dear Sir,
First of all thank you for your reply.
The scale consists of two factors and we will use factor totaal score instead of sumscore. The communality scoes are high enough.
You can not do this because you do not know any idea about prior pool of variables before extraction performed by the developer. Instead you should apply CFA for confirming the factors with that variables. Otherwise new/modified version of the scale measures different thing rather than developer offered.
@Sadri, you are right in principle, but having a scale being developed does not mean it can be generalized to all situations. That would be theoretically possible if you have a Rasch scale that does not show dif. I have not seen this often. I ran Mokken scale analysis on the MOS 36 years ago and could show the structure to be the same in our sample, but one item showed DIF and that is not very nice as it implies a difference in kind and not of quantity. If you want to test you should fix all parameters of the previous research and see if the model holds (a good fit and all other indicators of fit good as well (RMSR, AIC)). Using AMOS, LISREL or Mplus you could see if it is ok and also see where misfit exists.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Probability
- Probability Theory
- Reliability