I'm conducting a CFA on school data (ordinal variables), so I'd like to take clustering into account. Only part of the variables within the scales have substantial ICCs. Is it wise to fit a "full" factor model on the within-level and a "partial" model on the between-level, so the loadings of variables with low ICCs would be 0 on the between factors.
How would you interpret those findings and the loadings on the within and between factors?
bmuthen posted on Thursday, November 24, 2005 - 8:14 am
Yes, you can fit a 2-level factor model where the loading estimates will then most likely be small for the variables with small iccs. Note that you can have zero between residual variances and that is ok. The items with substantial between loadings then represent the school-level factor - for example, with achievement data you may find that only reading variables and not math variables measure school excellence.
So you would say, it's also a kind of model selection strategy to "exclude" variables from between level loadings because of small or neglectable ICCs?
Thank you very much fpr your comments.
bmuthen posted on Friday, November 25, 2005 - 8:22 am
Yes, you can do that type of exploration.
MReis posted on Tuesday, December 06, 2005 - 3:31 am
I have two related questions to the above topic: - The ICC is an index of the proportional amount of between-level-variance, but is there a rule of thumb about how many variance in absolut terms is "necessary" for successfull two-level-modeling (robust estimates, etc.)? - Regarding the discussed selection strategy of excluding items with low ICC/variance from the between-level: Are two models nested, that simply differ in the number of between-level-items (i.e. fixing their between-level-variance to zero) and therefore could be used in a log-diff-test?