joshua posted on Wednesday, July 07, 2010 - 12:38 pm
In a project that i am working on, we are comparing a shortened version of an instrument with four indicators of a proposed latent variable in a different population than what has been previously studied. Thus, I thought that CFA would be appropriate to test whether the previous conceptualization of this latent variable is appropriate for the population in question.
Because we had to shorten the measure, we also lost some internal consistency (alpha for each of the indicators .41,.50,.53, and 79), which could be due to alot of things (ability level, # of items, etc). When i run the CFA, the squared multiple correlations for these observed variables are (.33,.25, .31, .19).
My question is whether or not lower indicator reliability is problematic,and if so, how much of a problem is it, and at what values/cutoffs? Would the values stated above be considered problematic? The model fit is good (perhaps too good), CFI=1, RMSEA =.00, 90% LO/HI CI=.00,.00; TLI=1.226, SMSR =.003, but with such great fit, it makes me leery whether or not i did things right. Are there other outputs that I should be paying attention to evaluate the viability of constructing such a latent variable?
Would you recommend omitting indicators on basis of low individual loadings in an CFA, or would you consider it fairly arbitrary given that the same indicator could be assigned to be the anchoring indicator in another parametrization?