anonymous posted on Friday, February 17, 2006 - 9:48 am
Hi, Dr. Muthen,
I am wondering if my model is identifiable.
I have repeated measures on three constructs (A, B, and C). The repeated measurements are NOT structured at all. Patients have different number of measurements, different time points, and different range of time periods. I am fitting a linear growth model for each construct (with random intercept and random slope) and what I really want to do is to predict "slope of C" by "slope of A" and "slope of B". I had no problem up to this point.
Now I want to include an interaction between "slope of A" and "slope of B" to predict "slope of C". I created an interaction term using "xwith" command. I tried several variations of my model. None worked. I got either "NOT ENOUGH MEMORY SPACE ..." or "NON- POSITIVE FISHER INFORMATIOn MATIRX ... DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION ... "
I am wondering if the model with interaction between two slopes are even identifiable when the time points are so unique (no common values).
bmuthen posted on Friday, February 17, 2006 - 5:14 pm
It sounds like this model should be identified. Not enough memory space suggests that you have many dimensions of integration - you can check in the error message and also by requesting Tech8 and looking at the beginning of the screen printing. The interaction itself only contributes 1 more dimension so it sounds like you don't have continuous outcomes because if you don't you have many dimensions of integration even without the interaction.
The Fisher info matrix message does suggest non-identification, but you would have to send the input, output, data, and license number to firstname.lastname@example.org to get this figured out.
We are examining the effects of int and slp of 2 variables measured at 4 time points, as well as their interaction, on change (slopes) in mental health. We have tried using the xwith command to calculate our interaction terms between the two intercepts. However, we have many intergration points, and we are not sure why. Is it possible our model is saturated? or that the use of time scores in the LGM is causing problems? Any advice is appreciated. Thanks
I have a general question about interaction effect between two latent slopes. If I follow the traditional Aiken and West method , the simple effects of the independent variable (depending on the high versus low levels of the moderator) can be tested. If I find that two latent slopes interact, how can I test the simple effects using Aiken and West's method?
I have run a parallel process growth model. My main interest was an interaction between the two slope terms, which I created using XWITH. While I found the interaction to be significant, the estimates for several of the covariates in the model changed quite dramatically when adding the interaction. Most notably the baseline covariate of the outcome variable (correlated at .5) became non-significant. Based on what I have read on the Mplus support site, it seems that the model with latent interactions is a different model, so the estimates will change. However, I am wondering whether they are reliable to report, or whether it is only important to report whether or not the latent interaction is significant? Thank you in advance!