| Message/Author |
|
|
| Jian Wang posted on Thursday, January 29, 2009 - 3:31 pm
|
|
|
Dr. Muthen:
I am using Mplus to find the indirect effect between a binary outcome and a categorical predictor, with two categorical mediators. Therefore I tried to use the bootstrapping approach.
I understand that the estmites in the output are assessed using probit regression. I wonder if I am able to use logistic regression in bootstrapping? I tried to use "estimator is ML", which gave me error message.
Thank you for your help. |
|
|
|
|
| Bootstrapping is available for the ML estimator. Please send your Version 5.2 output and license number to support@statmodel.com. |
|
|
| Jian Wang posted on Thursday, January 29, 2009 - 3:42 pm
|
|
|
| Great thanks for your promp response. I am using the Demo version currently. Is this the reason I could not use ML estimator? |
|
|
|
|
| No, the demo version is the same as the regular version except for a limit on the number of variables. Please send your output to support@statmodel.com. |
|
|
| Ben Spycher posted on Wednesday, March 04, 2009 - 2:11 am
|
|
|
Dear Linda,
I am doing a simulation study in which I generate binary outcomes alternatively from factor models, latent class models and factor mixture models. I want to fit the generated data with the various models to see if the true stucture is recovered.
To speed up the process it is convenient to use the starting option of the Montecarlo command to read in starting values for the parameters. However the montecarlo command only seems to handle the situation when data generation and estimation are of the same model class. If I use external montecarlo (option montecarlo in the data command) there does not seem to be an option to read starting values from a file. Do you have a suggestion? |
|
|
|
|
| Your understanding is correct and I have no suggestion. The STARTING option is for only internal Monte Carlo. However, with external Monte Carlo you must give values in the MODEL command for each parameter. These are used for coverage and also as starting values. So I'm not sure why you would need the STARTING option. |
|
|
|
|
Thanks. The reason why this would be convenient is that for models with many parameters it is quite cumbersome to fix the starting values using the syntax of the model command. It would be much more convenient to read in values from an already fitted data set as starting values. As starting values potentially greatly increase convergence time, might this be a useful feature to include in the next version?
And/Or extend the Monte Carlo command to be able to generate from one model class and estimate from a completely different one?
Kind regards and thanks for your help
Ben |
|
|
|
|
hi, I'm running a mediation analysis which is moderated as well. I've got 2 groups and 3 measurement points. at point 3 I've got in one group 66 people and in the other just 9 left. now following warning shows up:
GROUP 2:
WARNING: THE SAMPLE CORRELATION OF PDAUER_3 AND E_2003 IS -1.000
DUE TO ONE OR MORE ZERO CELLS IN THEIR BIVARIATE TABLE.
INFORMATION FROM THESE VARIABLES CAN BE USED TO CREATE ONE NEW VARIABLE.
GROUP 2:
WARNING: THE SAMPLE CORRELATION OF PDAUER_3 AND K_VORH03 IS -0.986
DUE TO ONE OR MORE ZERO CELLS IN THEIR BIVARIATE TABLE.
INFORMATION FROM THESE VARIABLES CAN BE USED TO CREATE ONE NEW VARIABLE.
AND NO bootstrap draws are complteted anymore (I've requested 5000).
I guess the warning shows up and the bootstrap draws are missing because of the unequal group size?!
do you think I should at that point stop splitting up my sample into groups and just use all of the people (without grouping)?
or even more don't use groups right from the beginning?
Or could I do something else?
Hopefully you'll help me.
Best wishes. |
|
|
|
|
| The error messages you show are definitely the result of small sample sizes. I would think this is the cause of your other problems as well. |
|
|
| Back to top |
