| Message/Author |
|
|
|
|
Dear Professors,
I am wondering which is the best estimator to use for my data in a CFA. Some of my indicators for the latent factor are categorical and some are continuous. I was hoping to use ML to best account for missing data but when I run this with categorical variables it does not give me the CFI/TLI or RMSEA. What is the best way of assessing model fit in this situation? Or would you recommend I use WLSMV as the estimator instead?
Many thanks,
Virginia |
|
|
|
|
| If you want fit statistics like chi-square and related measures, you should use WLSMV. If you have a lot of missing data, you should use maximum likelihood. For each factor with categorical indicators and maximum likelhood, one dimension of integration is required. This is also true for residual covariances among categorical indicators. If you have more than four, you should consider weighted least squares. |
|
|
|
|
Hi Linda,
Thank you for your advice. Just to follow up - if one uses maximum likelihood (as I hope to) how should one judge the absolute fit of the model?
Many thanks,
Virginia |
|
|
|
|
| There are no absolute fit statistics when means, variances, and covariances are not sufficient statistics for model estimation. You can compare nested models using -2 times the loglikelihood difference which is distributed as chi-square or you can compare models with the same set of dependent variables using BIC. |
|
|
|
|
Hi Linda,
I understand. Thank you for your advice.
Virginia |
|
|
|
|
I am using MPlus for a CFA with a combination of binary and polytomous variables. I have about 700 observations and am running a 1 factor model. The distributions are as follows:
WORKUNST
Category 1 0.821 1213.677
Category 2 0.179 265.241
DEPRESSN
Category 1 0.317 468.190
Category 2 0.683 1010.728
DEPRESSI
Category 1 0.202 299.400
Category 2 0.798 1179.517
ILLNOMON
Category 1 0.038 56.263
Category 2 0.103 152.714
Category 3 0.271 401.880
Category 4 0.589 874.089
My question is - what is the best estimator to use for the CFA? I have read a lot on line and it seems that MLR is suggested but does not give the fit statistics I usually use (RMSEA, CFI, TLI). Is WLSML inappropriate? Or is there some other estimator you would recommend? Please advise. |
|
|
|
|
| The default in this case is WLSMV. You can use that or MLR. With MLR no absolute fit statistics are available. With WLSMV they are. If you have a lot of missing data, would recommend MLR. |
|
|
| Back to top |
