Treatment of code 6

2 replies [Last post]
wuhao_osu's picture
Offline
Joined: 09/07/2010

Hello,

I got several code 6 when running simulations. Some of those cases have a near singular estimated hessian, though the calculated hessian is fine. Is this the source of non-convergence? Would it work to re- run the data with the parameter estimates as initial values? Is it possible to have the optimizer use the calculated hessian all the time?

Thanks

- Hao

neale's picture
Offline
Joined: 07/31/2009
Yes, code 6 is more likely to

Yes, code 6 is more likely to occur if the estimated Hessian is inaccurate, especially if it is nearly singular.

Indeed it is a good idea to rerun such models from their solutions, which is extremely easy to do in OpenMx.

fittedModel <- mxRun(originalModel)
refittedModel <- mxRun(fittedModel)

This procedure may be repeated if refittedModel fits better than originalModel. If it doesn't then a third run is unlikely to help.

Also, it is a good idea to try other starting values in the originalModel

tbates's picture
Offline
Joined: 07/31/2009
sounds like good information

sounds like good information (in the "what to do now" section ) to add to the code 6 error message