Message/Author |
|
Anonymous posted on Wednesday, March 30, 2005 - 4:56 pm
|
|
|
I have a measure with ordinal items. As it measures illness and many respondents are not ill there are a number of the items which have a very low response rate at the most extreme (severe) category (e.g., 10/1500 responding '4' on a 1-4 scale). Is there a point at which one should consider collapsing categories? My specific analyses involve EFA/CFA, but any thoughts on the issue more generally would also be welcome. Thanks. |
|
BMuthen posted on Saturday, April 02, 2005 - 9:02 pm
|
|
|
See Muthen (1989) in the special issue of Sociological Methods & Research. It should be in the reference list on the website. |
|
|
Hi, I’m testing the factor structure of an instrument using exploratory and confirmatory factor analysis with categorical factor indicators in a sample of 700 partcipants. The instrument consists of 20 items which have to be answered on a 5-point scale. For most items low respons frequencies (N varies between 1 and 7) are observed with regard to the highest respons categories, and for one item the respons frequency on the highest respons category is zero. Even after reading the article regarding estimator choices with categorical outcomes on the MPlus website, I’m not sure which estimator is the best option in this specific situation. Would you advice in this situation (1) to use the default WLSMV estimator, (2) to use the ML/ MLR estimator, or (3) to merge the two highest respons categories into one and use the default WLSMV estimator? Or is there maybe a better alternative that I don’t see? Thank you very much for your reaction. |
|
|
The answer depends on how many factors you have. As the estimator note says, ML(R) is not suitable for say more than 3 factors. But MLR is probably better at handling sparse item categories. In either case I would collapse the 2 highest categories. |
|
|
Thank you for the quick respons. The hypothesized number of factors is 4. Will this be computationally too demanding? |
|
|
It is possible to use MLR with 4 dimensions - you can try it - but it is slow and may give low precision. You can use integration = montecarlo(5000); Low precision problems are seen if there are negative ABS changes in the TECH8 loglikelihood printing from the various iterations. |
|
Back to top |