I have a similar problem, and integration is definitely required due to the categorical variables. Using WLS avoids the problems but my parameter estimates are very different, which I will have trouble explaining. It's not unexpected it is a fairly different algorithm.
My thought is that for obtaining the indirect estimates then parametric bootstrapping would fix the problem, but this isn't available in Mplus either. One option seems to be to create my own simulated datasets and then run Mplus on these and calculate the bootstrap standard errors. Is there an easier or better way ?
Weighted least squares provides probit regression coefficient. Maximum likelihood provides logistic regression coefficients. You should be comparing column three of the output, the ratio of parameter estimates to standard errors.
You can use MODEL CONSTRAINT to define indirect effects.
Hello, I am estimating the model with the latent classes and each class is allowed to have different parameters in equations. From mplus I receive the following message "Some statements are only supported by ALGORITHM=INTEGRATION". Once I specify ALGORITHM=INTEGRATION, the program is running. But it says: the dimension of integration is zero and the total number of integration is one. Basically, the there is no integration. So, why do I need this integration? And, What is it does in my case? P.S: some of my variables are categorical and once I remove the command “categorical” the program runs but it doesn’t make sense any more.
I am trying to run a continuous-time survival analysis using the Cox regression model and I also want to test indirect effects whitin this model. However, since survival analyses need algorithm=integration, the model does not run with indirect effects.
Is there a way to test indirect effects when using continuous-time survival analysis using the Cox regression? What do I need to do if this is possible?
Thank you for reply. I would like to test two mediation paths. Both mediators have continuous scales. The final variable is dichotomous (no/yes).
I also tried to specify indirect effects with the contraint command, however than Mplus comes with a fatal error that montecarlo integration is needed for the analyses (?). Also, when I do not specify any mediation paths, Mplus still gives me a fatal error montecarlo integration is needed (I guess this error is related to the scales of the mediators?). Is there anything else we can do?
Thanks again for your reply. I indeed have missings within the mediators, however, I thought FIML would take care of it. Does FIML not work when using these kind of analyses? Would missing imputation be a better idea and afterwards try to run the model again with MODEL CONTRAINT to estimate indirect effects and INTEGRATION=MONTECARLO? Or is INTEGRATION=MONTECARLO not necesarry when the missing data is imputed?
Or would it be easier to run the model in two steps, and than calculate the mediation effect by hand using a Sobel test?
If your model requires numerical integration and you have missing data on the mediators, Monte Carlo integration is required. This does not mean FIML is not being used. If you want to use imputed data instead, that is an alternative. I do not see a need for this.
Running a model in two steps is never a good idea if it can be done in one step.
Apologies for coming back to you again. I am quite new to this and seem to run into difficulties as my outcome variable is categorical. I have done as you suggested but I am now getting the following error:
FATAL ERROR THIS MODEL CAN BE DONE ONLY WITH MONTECARLO INTEGRATION.
I am not sure how to proceed and would be grateful for any advice.
Apologies for coming back to you again about this. The more I read about the topic, the more I am unsure what I am doing is correct.
Could you tell me is it possible to include a multiple mediation effect in a multinomial logistic regression? I have a categorical (6 categories) outcome variable and four mediators (all continuous) and a continuous x variable.
I can find many references on mediation with binary outcome variables but not multinomial outcomes.
I have not seen a thorough statistical treatment of indirect effects when the Y variable is nominal (unordered categorical) using multinomial logistic regression. I would recommend simplifying to working with one binary Y at a time, where the binary outcome is formed as one category versus all the others.
Thank you for your response. I have now rerun my analysis using a binary outcome variable as suggested.
But now I am getting this error:
WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES. CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE CLASS.
I have checked TECH4 output but I am not sure what the problem is. I'm quite new to this but is there anything you can suggest to solve this?
The default for a count variable is maximum likelihood using numerical integration. You cannot treat the sum of a set of count variables as a count variable. If you remove the COUNT option and treat the variable as continuous or censored, MODEL INDIRECT will be available.
Jenny L. posted on Saturday, May 11, 2013 - 9:46 am
Thank you for your advice, Prof. Muthen. But I thought we can't treat count variables as continuous because they are not normally distributed? Even though the count variable I mentioned above is an average of 2 coders' counting results, the data distribution is still not normal. In this case, can I remove the COUNT option and treat it as a continuous variable?
The sum or average of a count variable is not a count variable. That is why I mentioned continuous or censored. Censored may be the better choice. But you can't treat the sum or average as a count variable.
Jenny L. posted on Saturday, May 11, 2013 - 11:47 am
Thank you for your suggestion, Prof. Muthen. I treated the mediator as a censored variable (cesored from below),but I still got the same error message:"MODEL INDIRECT is not available for analysis with ALGORITHM=INTEGRATION." Could you advise how I can fix the problem?
Here's the code I wrote:
CENSORED IS OwnP_T1(b);
ANALYSIS: TYPE IS missing; ESTIMATOR IS MLR; ITERATIONS = 1000; CONVERGENCE = 0.00005;
I had missing values in my independent variable and I added integration = montecarlo in the analysis command. Now I get this warning message: "The INTEGRATION option is not available with this analysis. INTEGRATION will be ignored. Specify ALGORITHM=INTEGRATION to use this option."