Bayesian SEM in AMOS: Waiting to accept a transition before beginning burn-in

[Yang, Hongwei]

Dear Friends,

 

I have a question regarding the Bayesian SEM platform offered in AMOS. I have this question after reading and following the SPSS tech-note found here: http://www-01.ibm.com/support/docview.wss?uid=swg21476869.

 

I am running the Bayesian analysis for a model that contains 7 to 9 dichotomous variables (Most of them are endogenous with only one being exogenous) and two categorical variables with one containing 3 categories and the other 5 categories. I treat all of them as ordered-categorical, although some of them seem to be just nominal categorical. I base my decision on another tech-note from SPSS: http://www-01.ibm.com/support/docview.wss?uid=swg21478651

 

 

I keep getting this warning message: Waiting to accept a transition before beginning burn-in. Based on the first tech-note above, the message seems to indicate that my analysis has kept obtaining parameter vectors that are either of low probability or un-defined. And I need to address this issue before being able to start the burn-in process. So far, I have tried the following:

1) All models in the analysis have passed the check under the maximum likelihood estimation, so they are all identified.

2) When using the Random Walk algorithm, the tuning parameter has been reduced to as low as .0001 through the use of the Adapt button.

3) The other MCMC algorithm (Hamiltonian Monte Carlo) has also been tentatively tried by manipulating the two parameters.

Despite the efforts, I still end up getting the above warning message. Is there anything else that I can try that can get the burn-in process to start?

 

My data set contains a small fraction of missing values on several of the variables. So, as usual, the AMOS starts with the imputation of missing values using the EM algorithm. Because there is some randomness in the imputed values, should I settle down with a fully imputed data set first and *separately* (say, using SPSS Statistics, instead of SPSS AMOS) before subjecting the fully imputed data to Bayesian estimation in SPSS AMOS?

 

Thank you!

 

Patrick Yang, University of Kentucky