Mark LaVenia posted on Wednesday, September 26, 2012 - 3:16 pm
I just downloaded V7 and am trying get my head around the many new features and how they might help me solve a modeling problem I've been having. From my read so far, I'm think the CROSSCLASSIFIED and BAYES commands will be key.
Using a 34 item, 4-category Likert-type observation protocol, 70 teachers were observed: 40 of the teachers were observed by a single observer and 30 of the teachers were observed by a pair of observers. The total 100 observations were distributed among 10 observers. Thus, teacher is crossed with observer.
I'd like to test the fit of various models (e.g., the 6-factor model suggested by theory vs. a 1-factor solution).
Any suggestion on what seems like the most promising approach would be greatly appreciated.
This sounds reasonable - you should run some univariate models first before you run cross-classified sem.
Mark LaVenia posted on Thursday, September 27, 2012 - 6:53 am
Thanks for your reply. I have run it--splitting the observer pairs, essentially creating two samples of 70 (40 observations common to both samples and 30 unique). On both samples, model fit indices exceed criteria for close fit.
For the crossed model, should I have items structured long, whereby subscale_ID and observer_ID would be the between clusters? If I structure items as wide, to have more than one between level, would I arbitrarily split the paired observations into between levels Obsrvn_1 or Obsrvn_2 (each containing a vector of IDs)? Alternatively, would I specify a between level for each of the 10 observers (using Obsvr1-Obsvr10 dummy variables as the clusters)?; all of the examples I've seen only have two between level and, thus, I don't know if this is even possible.
I've already looked at a number of resources, including the V7 user's guide and slides from your and Bengt's August presentation at Utrecht Univ. Maybe I missed it, but I didn't see specifics pertaining to the kind of data I have. Please direct me where I might look to get greater clarity on the particulars of modeling these data.
If it helps to understand my question: My goal is to be able to respecify the model as needed (collapse subscales, etc.) and extract factor scores that can be used in phase-2 analyses.
Thank you for your time and thoughtful suggestions.