Residual covariance categorical observed
Message/Author
 Jan Olav Christensen posted on Friday, September 11, 2015 - 1:36 pm
I have a question regarding a two-wave cross-lagged model with two latent variables with categorical indicators and one observed categorical variable.

I have read that categorical dependent variables should not be related to latent variables by WITH, but rather by regressing the latent variable on the observed categorical. However, in my model I cannot state that observed_categorical_dependent -> latent_continuous at time 2. If I specify it with a WITH statement, the estimates are different. Also, I have related factor indicators over time with WITH statements, although they are categorical.

1. Does it make sense to use WITH for residual covariance of a latent with an observed categorical?
2. Does it make sense to correlate observed categorical items over time with WITH?
3. Is it wrong to covary an exogenous observed with an exogenous latent using a WITH statement? It does not produce a warning or error statement.
 Bengt O. Muthen posted on Saturday, September 12, 2015 - 5:18 pm
The idea that that categorical variables should not be related to latent variables by WITH, but rather by regressing the latent variable on the observed categorical is only relevant if the categorical variable is exogenous. Otherwise, WITH is fine. I assume you use WLSMV. Then the answers are:

1. yes.

2. yes.

3. it is not wrong. It is just that with WLSMV is adds an assumption of underlying normality for the categorical exogenous variable.
 Jan Olav Christensen posted on Friday, October 02, 2015 - 4:18 pm
Dear Drs Muthen,

I am wondering about something that may be obvious to everyone else:

If you have

CATEGORICAL = x1-x3 y ;

f1 BY x1-x3 ;

y ON f1 ;

the regression of y on f1 is probit.

But what if you have

CATEGORICAL = x1-x6 y ;

f1 BY x1-x3 y ;
f2 BY x4-x6 ;

y ON f2 ;

Is y on f2 probit or linear?

 Bengt O. Muthen posted on Friday, October 02, 2015 - 6:09 pm
If you use ML you have a choice of link=probit and link=logit. There wouldn't be a different link for the two cases you mention.

With WLSMV and Bayes the link is probit. There wouldn't be a different link for the two cases you mention.

You can see in the Analysis Summary which link is used.
 Jan Olav Christensen posted on Friday, October 02, 2015 - 11:16 pm

So, a categorical factor indicator is not treated differently when it is on the lhs of an ON statement than it is on the rhs of a BY statement? What if what I want to do is to regress the residual variance of y on f2?
 Bengt O. Muthen posted on Sunday, October 04, 2015 - 4:54 pm
Q1. correct.

Q2. That can be done only with a continuous y.