Message/Author |
|
|
I am estimating a WLSMV model where two of the predictors are correlated. One of the predictors is age (with a mean of 25) and the other is a variable that has been centered. Within the model statement, I correlate these using the "WITH" statement but, in the output, the resulting estimate is greater than 1 (see below). Can you please shed light on what is happening? Estimates S.E. Est./S.E. Std StdYX 2.248 0.595 3.779 2.248 0.172 |
|
|
Means, variances, and covariances of observed exogenous variables should not be part of the estimated model. The model is estimated conditioned on these variables. You should remove the WITH statement. When you include it in the model, you make unnecessary distributional assumptions about these variables. The estimate you are receiving is a covariance not a correlation. |
|
|
Hello, I'm trying to replicate a comparison between fixed- and random-effects models in Mplus with a categorical dependent variable, WLSMV estimation with Theta parameterization. How might I specify that the covariance between an observed exogenous variable and the exogenous latent variable (the latent time-invariant variable) is 0 per the random effects specification? The two examples I've seen have continuous DVs and use ML estimation and specify "eta WITH IV@0" or "eta ON IV@0" with no particular issues (where "eta" is the latent time-invariant variable and "IV" is one of the exogenous predictors). However, when I do the same and try to run a difftest to check the chi-square difference between the fixed and random effect condition, the models are not nested. I think this has something to do with covariances among the exogenous variables. Can I specify covariances between the latent time-invariant variable and each observed exogenous variable in the fixed effects model, then set them to zero for random effects? Any thoughts you have would be most appreciated. Thanks, Dana |
|
|
I misspoke above about the nesting of the models: the "fixed effects" model isn't purely a fixed effect model. The issue is that in the examples I'm looking at, the IV is not part of the estimated model when "eta WITH IV@0" or "eta ON IV@0" is part of the model command. However, these models are ML with continuous dependent variables. When I do the same thing, but with a dichotomous DV and WLSMV, the IV becomes part of the estimated model (the mean is estimated), and I'm trying to figure out why. |
|
|
With a dichotomous DV and WLSMV, the IV should not become part of the estimated model when you say eta ON IV or eta ON IV@0. IV becomes part of the model when you include it in WITH or mean and variance statements. If this doesn't help, please send to Support. |
|
|
Thank you for your response. I see now that the default is that the latent time-invariant variable "eta" is not correlated with the predictors, which is why "eta on IV@0" did not change the model. If I want eta to be associated with other exogenous variables in the model, I need to specify "eta on IV" for each exogenous variable, then constrain to 0 to test whether they need to be associated. Does this seem right to you? Thanks again, Dana |
|
|
You can add them but I would not constrain the non-significant ones to zero. This may not be replicated in another sample. |
|
Back to top |