Dear all,
I have been running a model within a loop over several variables, which does not converge well for all variables. Therefore, I applied mxTryHard. There are two things I noticed:
1) for models that run well without calling mxTryHard the best fit does not seem to change using mxTryHard, but the status code changes from 0 to -1
> fit1 <- mxRun(model,intervals =F)
> fit1$output$Minus2LogLikelihood
[1] 4595.342
> fit1$output$status
$code
[1] 0
$status
[1] 0
fit2<-mxTryHard(model,intervals=F)
> fit2$output$Minus2LogLikelihood
[1] 4595.342
fit2$output$status
$code
[1] -1
$status
[1] 0
2) even when I set "intervals=F" (like in the example above) the mxTryHard option computes intervals.
Can you help?
Thanks in advance
Printer-friendly version
Thanks Rachel. The first bug you mention should already be resolved in the development version of OpenMx. The second is a good point, I will adjust this. These changes will be available in the next release, or by building OpenMx from source.
Thank you!
Hi all,
I experience the same status code = -1 problem in case of any model I run with mxTryHard, be it a model that results in "Solution found", or a model that results: "OpenMx status code 1 greater than 0; Not all eigenvalues of hessian are greater than 0; Retry limit reached, "
I'm running OpenMx version: 2.3.1 [GIT v2.3.1], R version: R version 3.2.1 (2015-06-18), Platform: x86_64-w64-mingw32, Default optimiser: NPSOL, installed from the website.
Does this mean both models converged (MxStatus Green)?
Status code = -1 means premature termination of the optimizer, so I don't think any models converged. Based on the information in your post, I'm not sure what's going on, though. Could you copy-paste the messages printed to the screen by
mxTryHard()?Thanks for having a look at it!
Limiting the extra runs to 1 to supress output (output the same for 10 extra runs), mxTryHard returns:
@output$status returns:
In case of a model where it says, solution found:
But the status is the same as the previuos model:
Hmm. OK, it would appear that the optimizer is NOT terminating prematurely. I would guess that the second model in your post converged, but that the first one (with just one extra try) didn't.
I think the status code of -1 is spurious, and due to a bug.
Thanks for the information! mxTryHard really seems like a nice tool, hope the bug can be fixed easily. Is there a way to identify mxTryHard runs which terminate with a solution, other than the status code for the time being?
Two other interesting things, which might be bugs is:
-if I rerun a fitted model, which was fitted with mxTryHard with the simple mxRun command, the status code stays -1 for all models.
-if I don't include the "intervals = T" command in the mxTryHard command, and rerun the fitted model with the simple mxRun command with the "intervals = T" command, then it doesn't calculate the confidence intervals.
Thanks for pointing these out, they all stem from the one issue which needs to be fixed really. For the moment, a work around would be to use omxGetParameters() to extract the estimates from the mxTryhard with intervals=F solution, and omxSetParameters() to assign these values to an unfitted model, then fit this with mxRun and intervals=T. Or just pay attention to whether mxtryhard reports 'solution found' or not - this indicates that the fit prior to estimating confidence intervals was accepted, and confidence interval estimation won't change this. Sorry for the fuss! :)
deleted my comment and taking a look