DTSA Power Analysis PreviousNext
Mplus Discussion > Growth Modeling of Longitudinal Data >
Message/Author
 Christian M. Connell posted on Tuesday, October 20, 2009 - 8:56 pm
Can the monte carlo utilities be used to estimate power for covariate effect in a DTSA model? I was trying to use the corresponding monte carlo code for example 6.20 as a guide to develop a model to estimate survival across 7 periods, but having some difficulty.

Specifically:
1) not clear exactly how to modify the population statement to set-up a given hazard model with low frequency event
2) the %sig for x on f (at least as I specified the model) jumps from 0.000 at one sample size (e.g., 225) to the next increment (e.g., 226).

Thanks for any guidance you can offer!
 Christian M. Connell posted on Wednesday, October 21, 2009 - 8:49 am
Sorry, just realized that #2 (above) should have read: he %sig for x on f (at least as I specified the model) jumps from 0.000 at one sample size (e.g., 225) to 1.000 at the next increment (e.g., 226).
 Linda K. Muthen posted on Wednesday, October 21, 2009 - 9:25 am
1. The thresholds are the parameters that control this. Higher thresholds represent more rare events.
2. Please send the two outputs and your license number to support@statmodel.com.
 Christian M. Connell posted on Wednesday, October 21, 2009 - 9:49 am
Will do --

I discovered the error that was leading to my sample size finding (had another model statement later in the code that was incorrect). However, now have an issue where the covariate effect power jumps the same way (0.000 at .3 vs. 1.000 at .31) for a given n. Will send outputs/license as indicated.

Thank you
 Pierre Walthery posted on Tuesday, October 14, 2014 - 11:13 am
Hello,

I am trying to do a post-hoc power analysis of a LGC model with categorical outcomes and covariates. The power analysis is about one regression coefficients on the slope factor.

1. I understand that the SS method does not apply to categorical outcome. Is that correct?

2. When I try to use the MCMC methods as in Example 12.6, I get an error message for each categorical outcome.

*** ERROR in MONTECARLO command
A CATEGORICAL variable in the analysis must be generated as a categorical
variable.

It looks as if MPlus think I am trying to run a simulation study. Am I doing anything incorrectly?

Thank you!


MONTECARLO:
NAMES ARE nql2 nxql2 nyql2 nzql2 x1-x10;
CATEGORICAL IS nql2 nxql2 nyql2 nzql2 ;
NOBSERVATIONS = 12000;
NREPS = 500;
SEED = 45335;
POPULATION = EST1.DAT;
COVERAGE = EST1.DAT;


MODEL POPULATION :
iw sw | nql2@0 nxql2@1 nyql2* nzql2@3;
iw sw ON x1-x9;
sw on x10;

MODEL :

iw sw | nql2@0 nxql2@1 nyql2* nzql2@3;
iw sw ON x1-x9;
sw on x10;

OUTPUT: TECH9 ;
 Bengt O. Muthen posted on Tuesday, October 14, 2014 - 6:03 pm
1. By SS method, I assume you refer to the population covariance structure approach - that does not apply to categorical outcomes.

2. Chapter 12 doesn't cover MCMC (i.e. Bayesian analysis), but Monte Carlo simulations using various estimators (you can go to a UG example with growth modeling of categorical outcomes and then find the corresponding Monte Carlo setup on our website to see how that is done). If you want MCMC estimation you have to look for example with Estimator = Bayes.
 Pierre Walthery posted on Wednesday, October 15, 2014 - 3:31 am
Thank you for your quick reply and apologies for the confusion: I indeed meant Monte Carlo simulation, not Markov Chain Monte Carlo.

To clarify: I have fitted a rather basic LGC model with categorical outcomes and covariates using robust maximum likelihood, exported the estimates into EST1.DAT. and used these as population values in a Monte Carlo simulation to get estimates of the power of the regression coefficient for x10, following the broad logic of example 12.6.

I managed to make it successfully work with another LGC model, which had continuous outcomes, but I get these error messages when I declare the outcome variable to be categorical as you can see above.

Does that mean that such a 2 stage MC simulation study cannot run with categorical outcomes and covariates? If so, are there any alternatives for me to estimate the power of this regression coefficient?

Thank you again.
 Bengt O. Muthen posted on Wednesday, October 15, 2014 - 2:50 pm
It sounds like you want to do Step 2, not Step 1 of UG ex 12.6. In which case you should look at the input on page 430. Here you see Data Type=MonteCarlo, not MonteCarlo: as you have it. And Step 2 does not have the Model Population statements.

If this doesn't help, send to Support with license number.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: