Anonymous posted on Wednesday, June 01, 2005 - 4:11 pm
I am attempting to use MPLUS to model the association between maltreatment, executive function and social dishinhibition in two groups of children: a foster care sample and an age and SES-matched non-foster care comparison sample. One of the inclusion criteria for the comparison group was that they have no recorded incidents of maltreatment. Thus, for the maltreatment variables, all of these children have a value of "0". (The foster children may also have a value of "0".) This is difficult because it confounds group membership with maltreatment status. Is it possible to use the KNOWNCLASS command to specify the maltreatment variable and thus account for the confound in the model? If not, could you suggest other ways to analyze the data using both groups?
bmuthen posted on Wednesday, June 01, 2005 - 11:53 pm
As a first step, let me ask you this:
1. With x, z, y representing maltreatment, executive fcn, and social disinhib, what is the model? x -> z -> y?
2. Is maltreatment a 0/1 variable or are there degrees of maltreatment?
Anonymous posted on Thursday, June 02, 2005 - 3:21 pm
Thank you for your reply. The answers to your questions are:
1) Yes. That is the model.
2) There are degrees of maltreatment. The particular score that I would like to utilize is a mean severity score that can range from 0 to 5. It is possible for the foster children to have a zero value on this score and all of the comparison children will be scored "0". (Alternatively, there are scores that I can use on which the comparison children will all have "0" and the foster children will have a range of values from 1 to 5.)
Again, thank you for your assistance with this!
bmuthen posted on Thursday, June 02, 2005 - 6:36 pm
Looks a little difficult to work with. Using the x, z, y notation earlier, I am not sure which group differences you are interested in, for the x->z relationship or for z->y? It seems like the x->z relationship can only be compared across the 2 groups by essentially estimating it in the foster group where x has variation, and then use that regression to predict z at x=0 in the non-foster group to see if this predicted z value is close to the observed z. The z-> y relationship may be easier to compare across groups. In any case, multi-group or knownclass could be useful. I was trying to think if the Mplus censored-inflated approach would work for x censored at x=0, but it seems like it would complicate things.
David Bard posted on Tuesday, March 20, 2007 - 3:08 pm
I'm trying a monte carlo simulation for a twolevel model with 3 known classes. How do I stipulate the known proportions in each class? I've tried creating a categorical group variable, specifying the logit intercepts in the overall model, and then fixing the intercepts in each class as shown below. This approach gets an error message: A categorical variable (generated as categorical) must be analyzed as CATEGORICAL.
%overall% %within% ... %between% [group$1*-.85];!30% of sample in class1 [group$2*.85];!30% of sample in class3 ...