Your patmiss and patprobs subcommands are the issue I think. The missing data pattern probabilities need to add up to 1. You should be able to figure it out by taking a look at the Mplus Users Guide sections describing these subcommands. Also, my free sample size planning course on goquantfish.com has an Mplus simulation example in which these commands are used to study a planned missing design.

Hi Christian Geiser

I studied the Mplus User's Guide to the best of my ability (which is limited when it comes to advance statistics) and watched your insightful tutorial.

I did not expect that PATMISS and PATPROBS might be the issue. They were. I assumed by saying that 50 % of all cases (.5) have a missing pattern of 5 % (.05) it would automatically be clear that the other 50 % have no missings at all... I have now specified two more PATPROBS with .3 and .2 and PATMISS with 7 % and 10 % missings, respectively. They now add up to one. The error message is gone.

Furthermore, I forgot to specify the thresholds, which is a painstaking undertaking. I set all of them to 0, which, in my mind, corresponds to an average probability of transitioning from one value to another on the categorical indicators. Is this correct?

Unfortunately, I know receive a less specific error message:

*** FATAL ERROR

THE POPULATION COVARIANCE MATRIX THAT YOU GAVE

AS INPUT IS NOT POSITIVE DEFINITE AS IT SHOULD BE.

I assume something is either missing in the MODEL POPULATION command or ill-specified. Sadly, mostly if not all literature out there only ever deals with continuous indicators...

I have attached a revised input file. If you find the time, perhaps you will be able to give me another hint as to what I need to change.

Many thanks

Marcel

Hi Marcel Grieger

It can be tricky to specify a population model from scratch in such a way that the population covariance matrix is positive definite and all population parameters make sense. One way to make it easier is to use a real data application as a template, save the parameter estimates from the actual application in Mplus using the SAVEDATA: ESTIMATES = ; command, and then running the simulation based on the estimates obtained in the real data application. See Example 12.7 (Step 1 and 2) at https://www.statmodel.com/usersguide/chapter12.shtml.

Provided the real application runs without problems, the underlying covariance matrix should be positive definite. This also has the advantage that your simulated parameter values are realistic and less artificial as they are based on an actual application.

Hi Christian Geiser

thanks again for coming back to me on my questions. I totally agree, it is really difficult for a non-statistician like me. I will definitely try the approach you suggested, however, does that not imply having already collected data? I thought the idea of a Monte Carlo simulation was to find out what N one needs in order to detect certain effects and to ensure good model fit before collecting data, hence the simulated population model.

If I follow your recommendation, would that be an a-posteriori Monte Carlo simulation, whereby I would be able to determine if I had enough power or whether I would need to get more data?

Questions upon questions...

Marcel

P.S. I e-mailed my input to the Mplus support as well. Let us see if they can work it out.

It is not necessarily a post hoc analysis. The idea here is that you may have some pilot data or data from a previous study/application that is similar enough to where those data could inform your simulation in a meaningful way. Otherwise, it may be difficult to pick meaningful population values for a simulation--the values are often rather artificially chosen. For example, in your simulation, all loadings were uniformly set at 0.1. This may not be a realistic scenario. Using a real data application as a "template" can make the population values more meaningful and help you find a positive-definite population covariance structure that can be simulated.

Hi Christian Geiser

that makes a lot of sense because defining all those values seemed fairly arbitrary to me. For example, how do I know what value I should assign to the second threshold of the fourth variable? Having some data to fall back on seems advantageous. Still, I would like to be able to set up a population model without any real data, but maybe this is just beyond my abilities. For the moment.

Anyway, I would not even have made it this far without your help. Thanks a million!