Clio Berry posted on Thursday, July 12, 2012 - 3:33 am
I'm new to SEM (and factor analysis) and am trying to work my way through creating an exploratory SEM model. I'm working with a 4 factor/latent variables strcuture with 4-6 indicators each.
I was working with WLSMV and had conducted the EFA, respecified in a CFA framework and was then adding my observed predictors into a full ESEM model (i.e. latent variables regressed on various predictors).
I have one variables which is ordered categorical but has 9 categories, so I believe it could be treated as continuous or ordered categorical.
I then read that MLR is recommended for non-normal data rather than WLSMV, and my data is non-normally distributed, but MLR won't converge with more than 3 factors when I attempt it for the EFA portion of my analysis.
If I run my EFA by CFA model (built using the WLSMV EFA) using MLR instead of WLSMV, then the fit statistics are much poorer, e.g. CFI .880 instead of .921, RMSEA of .064 instead of .045.
Are the 4-6 indicators for each factor continuous or categorical? What is the role of the variable with 9 categories.
Clio Berry posted on Friday, July 13, 2012 - 1:35 am
The variable with 9 categories is a measure of weekly hours spent in occupational activity (e.g. 0-5 hours, 5.5 - 10 hours and so on). The categories are not all equal to each other.
All the other indicators are continuous (mean scores from questionnaire subscales, plus one variable which is a 10 item global happiness score (0-10)).
The 9 category variable loads onto one (loadings of .3/,4) of the 4 latent variables. This variable loads onto the latent variable with another 3 continuous variables that all seem to refer to occupation, e.g. occupational satisfaction.
I am trying to estimate an ESEM model. I have two latent variables from binary factor indicators. I would like to regress these latent variables on a binary observed variable (gender) and a non-normally distributed continuous observed variable (age). As I understand, WLSMV is not robust to violations of non-normality of continuous variables. However, it also appears that I cannot specify variables as categorical under MLR in ESEM. In this case, would the best choice be to transform the non-normal continuous variable? Is there another solution that you can recommend?