|
|
 |
| Auxiliary models with BCH weights |
 
|
| Message/Author |
|
|
|
|
Hello,
I used Mplus to conduct a latent profile analysis and am now using BCH weights to test a regression auxiliary model combined with latent class regression. I have two questions.
1. Is it possible to do this type of analysis without using listwise deletion? I am losing a large number of observations.
2. I used model constraint and model test to evaluate whether profile means are significantly different from each other. I got the following error:
WALD'S TEST COULD NOT BE COMPUTED BECAUSE OF A SINGULAR COVARIANCE MATRIX.
However, I still got p-values for the new/additional parameters, indicating that some mean differences were significantly different from 0. Can you explain what that error message means and whether I can still trust the pairwise p-values?
Thanks so much! |
|
|
|
|
1. We do not use listwise deletion. There is a listwise command in Mplus but it is never the default. If you have a single dependent variable Y and some covariates X, the model for [Y|X] simply has an empty likelihood when Y is missing. Data containing information regarding [Y|X] is not removed.
2. Yes you can use the pairwise p-values.
For the Wald test the issue is the following. If you have 3 classes do not use
model tests:
0=m1-m2;
0=m1-m3;
0=m2-m3;
Instead use
model tests:
0=m1-m2;
0=m1-m3;
The third equation is redundant and causes singularity. |
|
|
|
|
Thank you so much, this is really helpful! (And I apologize for my delay in getting back to you).
I do have missing Y variables. Are there any other options for handling missing data? For example, is this compatible with multiple imputation?
Thank you again! |
|
|
|
|
| You can use multiple imputations (manually where each imputed data set is analyzed separately and then manually combine the results) however I would not recommend it unless you have some kind of a set of background variables that you can use for the imputation but you are not using in the model. If you don't have such a set of variables I would not recommend multiple imputation and our default method is the better option, which already uses all the information in the data, i.e., it is full-information FIML. |
|
|
| Back to top |
|
|
