In a multiple group analysis model, I was tring to use | and xwith to specify an interaction term of an ordinal factor (F, with orderdinal items) and an observed continuous variable (X). The outcome is also an ordinal factor (Y). Nevertheless, the programme did not run.
I was wondering if it is because | xwith is not available for multiple group analysis?
Should I instead manually form the interaction factor by creating "interaction items" using cross product of each item of the factor (F) and the observed variable (X)?
To do this, you need to use the mixture track and the KNOWNCLASS option. Following is a partial input:
VARIABLE: NAMES ARE y1-y6 x1-x3 g; USEVARIABLES = y1-y6 ; CLASSES = c (2); KNOWNCLASS IS c (g = 1 g = 2); MODEL: %OVERALL% f1 BY y1-y2; f2 BY y3 y4; f3 BY y5 y6; int | f1 XWITH f2; f3 on f1 f2 int; %c#1% f3 ON f1 f2 int; ANALYSIS: TYPE=MIXTURE RANDOM; ALGORITHM=INTEGRATION;
Thank you very much for your help. Can I ask if this set up is OK for multiple group analysis with continuous latent factor, ordinal categorical item please, or it is only for mixture model with multiple groups? The programme complained about the correlated residuals (repeated items). I took them out. But still programme did not run, says it had not enough memory space. Thanks! Kate
This is the way you need to do multiple group analysis if you have a latent variable interaction and categorical factor indicators. When you use the mixture track and the KNOWNCLASSES option, this is identical to doing multiple group analysis because class membership is known not unobserved. This is available for categorical outcomes only via the mixture track because numerical integration is required. With numerical integration each factor with categorical indicators and each residual correlation require one dimension of integration. We recommend no more than four dimensions of integration. See Numerical Integration in the user's guide for further information.
The programme says that there are 6 dimensions and 0.11391E+08 INTEGRATION POINTS. I guess that's the reason it did not run. Is this beyond the possiblity to run the model? What would you suggest to do please?
To speed up the computations, I hope you made sure to use
PROCESSORS = x;
where x is the maximum number of processors on your computer. For these types of heavy analyses you should have at least 4. Also, by requesting TECH8 you can see how long each iteration takes and get a reasonable prediction of how long it will take to get convergence (usually within 100 iterations). Look to see that you don't have a large number of negative ABS changes in the loglikelihood because if you do, this can be a sign that you don't have enough numerical precision in your integration. Also, I hope that you have first done a run with just 1 iteration, printing Tech1 to make sure you are considering the model you want before waiting for days.
Hans Leto posted on Thursday, May 17, 2012 - 2:09 pm
I am performing a multiple group analysis with a categorical variable. I would like to test its effect on the interaction (XWITH) between three continuous variables. I followed the syntax provided above in this forum (May 06, 2011 - 5:36 pm).
I have three-way interactions, when I perform it when a single interaction, everything is perfect, however when I include another interaction it gives me the message "ONE OR MORE PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE INFORMATION MATRIX."
In the input you provided how would you specify another interaction effect?
jas229 posted on Wednesday, June 06, 2012 - 2:28 pm
I am trying to run a multiple group comparison for a model that involves latent variable interactions. I am using the latent class approach that was described above. I tried to run one model where all of the parameters were constrained to be equal across the two classes and another model where all parameters were estimated freely across the two classes. I could run the model with parameters constrained to be equal across the two classes, but I was not able to run a model where all parameters were estimated freely. I received error messages about not declaring random effects variables in the models other than the overall model. However, I was able to run a model where structural paths (but not factor loadings) were freely estimated across the two classes. When I tried to compare these two models with a log-likelihood test, I got a nonsensical chi-square value. I also received error messages saying that ONE OR MORE PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE INFORMATION MATRIX. THE SINGULARITY IS MOST LIKELY BECAUSE THE MODEL IS NOT IDENTIFIED, OR BECAUSE OF EMPTY CELLS IN THE JOINT DISTRIBUTION OF THE CATEGORICAL VARIABLES IN THE MODEL. Based on some prior comments on the discussion board, I tried to run these models again with a higher number of random starts, but the model did not converge. I was wondering if you could provide some guidance as to how to proceed with this analysis. Thank you very much for your time.