|
Message/Author |
|
|
Hello Dr. Muthen, I have clustered data, with about 250 participants clustered within 6 teams. I'm interested in comparing a series of CFA models while also controlling for the clustered nature of my data. I understand that, with only 6 clusters, TYPE IS COMPLEX is not recommended. However, when I use a "fixed-effects" approach, i.e., by creating 5 dummy variables and adding them to my model as covariates, the CFIs and TLIs are much worse compared to the original (i.e., ignoring clustering) CFIs and TLIs. The RMSEAs remain similar, however. Why are the CFIs/TLIs so different across the models with/without the dummy variables? Which approach should I favour? Best regards, Ann-Renee |
|
|
I assume that your 5 dummy vbles influence only the factors of your CFA, which means that you specify full scalar measurement invariance across the 6 teams. That could be the cause of the misfit. When you ignore the grouping you are not addressing that issue. Do an invariance analysis using a 6-group analysis (I assume the 6 samples are independent): configural, metric, scalar. This can be done in one run using those Model= settings. |
|
Margarita posted on Tuesday, February 27, 2018 - 9:06 am
|
|
|
Hi Dr. Muthen, I understand that large cluster sizes are needed for clustering and multilevel modeling but out of curiosity I ran a 4-factor higher-order CFA with Type = complex to account for the clustering due to 5 countries (average cluster size = 197) which led to a substantially improved fit (ICCs range between .01 and .22): clustering: chi2 (295) = 499.47; RMSEA = .027 (.022-.030); CFI = 953; TLI = .948 no clustering: chi2 (295) = 2773.51; RMSEA = .092 (.089-.095); CFI = 869; TLI = .855 Is such an improvement something you commonly see when type = complex is used (irrespective of cluster size) and is it not be trusted due to the small N? Thank you for your time |
|
|
With only 5 clusters you should not trust these results. Instead, represent the clusters by 4 dummy covariates. |
|
|
With only 5 clusters you should not trust these results. Instead, represent the clusters by 4 dummy covariates. |
|
Margarita posted on Friday, March 02, 2018 - 1:10 am
|
|
|
Thank you Dr Muthen, will do that. |
|
Back to top |
|
|