You are generating variables as continuous and analyzing them as categorical. This cannot be done. The latest update will give an error message in this situation. It will be available today or tomorrow. Please don't paste output in Mplus Discussion as it may be large. Instead send the problem to email@example.com.
Anonymous posted on Sunday, September 19, 2004 - 12:55 am
When analyzing montecarlo results, Can the standardized parameter values be saved? It seems that the SAVEDATA command only saves the unstandardized parameters.
You can do this using the estimated parameters. The formulas can be found in Appendix 3. The technical appendices can be found on the website.
Daniel posted on Wednesday, December 15, 2004 - 2:35 pm
Hi Linda and Bengt, I'm having trouble running a simulation with three associated processes, two of which are continuous and one categorical. Is there something special I should know about this type of analysis?
Sorry. I haven't had time to take a close. I think that the problem is the population values that you have chosen result in no variance for the smoke varialbes, that is, all observations end up in one category. Where do the population values for the thresholds, for example, come from?
Daniel posted on Friday, December 17, 2004 - 10:49 am
I'll take a look at it again. I didn't change my population values to reflect the categorical nature of the variable. I ran this model first with all continuous variables. Then, I added thresholds values in a simple increasing (1,2,3) pattern, just to see if this model would work. In other words, I didn't take much effort into generating the values. I'll try again with better population values to see if it runs.
Daniel posted on Friday, December 17, 2004 - 11:06 am
Linda, I ran single process models for smoking (categorical) and BMI (continuous), and each ran fine. However, when I combined the categorical and continuous processes, I did not get any completions. So, I must be doing something wrong when I combine the two types.
With extreme thresholds of 2 and 3, when you combine the continuous and categorical outcomes, the correlations between continuous and categorical outcomes likely become hard to estimate. Use threshold values that are based on theory or real data and you should be fine. Remember that thresholds are z-scores. So a value of 3 would result in .13 percent in a category. With 1000 observations, this is about one individual.
Daniel posted on Friday, December 17, 2004 - 2:34 pm
Thank you very much, I'll try it
alexandra posted on Tuesday, February 22, 2005 - 9:47 am
I want to make a Montecarlo study with categorical variables but I don't know how to fix the paramters of the population's model. Indeed, I know the population's proportions so I am able to fix the thresholds but I don't know how to fix the path's coefficient, I would like there value to be meaningful and realistic but I only know the poluation probabilities.
I assume that you know the proportions because you have data that you have analyzed and have obtained the proportions from the data. See Example 11.7 in the Mplus User's Guide. Here real data are analyzed to obtain population parameter values for a Monte Carlo study. This may be helpful to you.
Jon Elhai posted on Monday, October 06, 2008 - 2:18 pm
I've been a bit unclear about what estimates to provide in a Monte Carlo study to estimate sample size and power. I seem to think that I should provide the unstandardized parameter estimates? But many of the Mplus Monte Carlo input files seem to center the predictor variables, making me think that standardized estimates would be okay.
Yo In'nami posted on Sunday, May 29, 2011 - 8:10 am
Regarding your comment right above, what would be the consequences of using standardized, not unstandardized, parameter estimates in conducting a Monte Carlo simulation study to determine power and sample size? Would we get overestimated or underestimated power values? I haven't been able to find references on this issue.
I am conducting power analyses on published SEM models, where authors reported standardized estimates only and rarely reported unstandardized ones except for multigroup analyses.
You would have to use MODEL CONSTRAINT to compute the standardized coefficients. Then the power would be correct. You can only get power for a parameter that is estimated. Standardized parameters are computed after model estimation unless you use MODEL CONSTRAINT.
You could consider doing a Monte Carlo study where all variables have variance one and use the standardized coefficients from the study as population parameter values. This would probably be close enough.
I'm having difficulty generating a categorical variable with a frequency distribution of my choosing. For example, I'd like to create a categorical variable with 5 levels, where the probabilities of levels 1,2,3,4,5 being "selected" are .2,.3,.1,.35,.05 (or really anything to this effect, the numbers I just listed were chosen arbitrarily). What I currently have is montecarlo: names=v6-v8; generate=v6-v8(4); categorical=v6-v8; where I would then (in model population) have to specify the thresholds for each of the levels. For example: [v6$1@somenumber v6$2@someothernumber v6$3@etc v6$4@etcetc]; I read in the Mplus user's guide that the data for categorical variables might be generated via a logistic model, so I tried converting my probabilities to that. I haven't been able to get that to work because I don't know the parameters of the underlying distribution from which the variables are generated. Do I need to specify the parameters of this logistic cdf, or have I made an error somewhere else along the way?