Preliminary analysis using latent variables

[Anna Warren]

Hello,

I am running a mediation model where I plan to use two latent variables as mediators. For preliminary analysis I want to show how these latent variables are related to my predictor/outcome (both continuous) (e.g. through a simple regression), but I am unsure how to do this.

The code for one of the latent variables is

sem (ypb8002 ypb8003 ypb8006 ypb8007 ypb8000 ypb8001 ypb8004 ypb8005 ypb8052 ypb8053 ypb8056 ypb8057 ypb8050 ypb8051 ypb8054 ypb8055<-Self), latent(Self) method(mlmv) stand cov(e.ypb8000*e.ypb8001 e.ypb8000*e.ypb8002 e.ypb8000*e.ypb8003 e.ypb8000*e.ypb8050 e.ypb8000*e.ypb8051 e.ypb8001*e.ypb8002 e.ypb8001*e.ypb8004 e.ypb8001*e.ypb8005 e.ypb8001*e.ypb8051 e.ypb8001*e.ypb8052 e.ypb8001*e.ypb8054 e.ypb8002*e.ypb8003 e.ypb8002*e.ypb8004 e.ypb8002*e.ypb8005 e.ypb8002*e.ypb8006 e.ypb8002*e.ypb8007 e.ypb8002*e.ypb8051 e.ypb8002*e.ypb8052 e.ypb8002*e.ypb8053 e.ypb8002*e.ypb8054 e.ypb8002*e.ypb8055 e.ypb8002*e.ypb8056 e.ypb8003*e.ypb8053 e.ypb8004*e.ypb8005 e.ypb8004*e.ypb8006 e.ypb8004*e.ypb8007 e.ypb8004*e.ypb8051 e.ypb8004*e.ypb8052 e.ypb8004*e.ypb8054 e.ypb8004*e.ypb8055 e.ypb8004*e.ypb8056 e.ypb8004*e.ypb8057 e.ypb8005*e.ypb8006 e.ypb8005*e.ypb8007 e.ypb8005*e.ypb8052 e.ypb8005*e.ypb8054 e.ypb8005*e.ypb8055 e.ypb8005*e.ypb8056 e.ypb8005*e.ypb8057 e.ypb8006*e.ypb8007 e.ypb8006*e.ypb8052 e.ypb8006*e.ypb8054 e.ypb8006*e.ypb8055 e.ypb8006*e.ypb8056 e.ypb8006*e.ypb8057 e.ypb8007*e.ypb8052 e.ypb8007*e.ypb8053 e.ypb8007*e.ypb8054 e.ypb8007*e.ypb8055 e.ypb8007*e.ypb8056 e.ypb8007*e.ypb8057 e.ypb8050*e.ypb8051 e.ypb8051*e.ypb8052 e.ypb8051*e.ypb8054 e.ypb8051*e.ypb8055 e.ypb8052*e.ypb8053 e.ypb8052*e.ypb8054 e.ypb8052*e.ypb8055 e.ypb8052*e.ypb8056 e.ypb8052*e.ypb8057 e.ypb8053*e.ypb8054 e.ypb8053*e.ypb8055 e.ypb8053*e.ypb8056 e.ypb8053*e.ypb8057 e.ypb8054*e.ypb8055 e.ypb8054*e.ypb8056 e.ypb8054*e.ypb8057 e.ypb8055*e.ypb8056 e.ypb8055*e.ypb8057 e.ypb8056*e.ypb8057)

I tried to generate a factor score for a latent variable to use in a regression, by using 'predict latent' but I get the following output:

. predict latent

output:
(xb(ypb8002 ypb8003 ypb8006 ypb8007 ypb8000 ypb8001 ypb8004 ypb8005 ypb8052 ypb8053 ypb8056 ypb8057 ypb8050 ypb8051
ypb8054 ypb8055) assumed)
too few variables specified

Is there a different way to do this or something I am missing?

Thanks

[Weiwen Ng]

Anna,

A first thought is that instead of manually specifying all the error covariances, you could use the covariance structure option to say that all your error covariances have an unstructured covariance, e.g.

sem (ypb8002 ypb8003 ypb8006 ypb8007 ypb8000 ypb8001 ypb8004 ypb8005 ypb8052 ypb8053 ypb8056 ypb8057 ypb8050 ypb8051 ypb8054 ypb8055<-Self), latent(Self) method(mlmv) stand covstructure(e._OEn, unstructured)

Normally, I think the error covariance structure would be diagonal, i.e. the variance of each error is estimated freely, but their covariances are constrained at zero. Unstructured means every covariance is freely estimated.

Speaking of that, is that a sensible assumption? Normally, we would assume that the latent variable is the only notable source of the covariances among the observed indicators. By my understanding, modeling a separate error covariance is done when you think there may be "excess" correlation among some indicators. Modeling excess correlation among all the indicators would seem to be excessive.

Finally, to your question, your syntax should have been:

predict self_predicted, latent

Without specifying latent as an option, -sem- thought you meant to generate a variable called "latent", and it assumed that this variable was the linear predictor.