gradient

[Carl Falk]

I'm wondering if there's any way to obtain the gradient (first order derivatives of the marginal log-likelihood) from a mirt model? In flexMIRT I believe an estimate appears in the -dbg file, if requested.

From mirt, I can see an estimate of the information matrix (with whatever standard error approximation was used) e.g.,

data <- expand.table(LSAT7)
(mod2 <- mirt(data, 1, SE = TRUE)) #standard errors with crossprod method
mod2@information

But, unless I've missed something, no gradient.

Thank you!

[Phil Chalmers]

Only the complete-data gradient require in the MML implementations, and it should necessarily be a vector of 0s at convergence, so it isn't returned. Currently the observed-data gradient is not computed, mostly because I don't really think it should be required if the EM converges in which case it should be a vector of 0's anyway. Even if the solution lands on a local min the observed-data gradient should be a vector of 0's.

Are you trying to extract it for a particular purpose (say, seeing what the observed-data gradient would be at some sub-optimal solution)? 

Phil

--
You received this message because you are subscribed to the Google Groups "mirt-package" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mirt-package...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Carl Falk]

Thanks for the super fast reply.

Precisely. The other catch is that I'm using custom item models as well (some currently not written in a general enough way for multidimensional models), with priors, that aren't yet in other software. At the moment just exploring mirt (and OpenMx) to see if it makes sense me me to use for research purposes instead of continuing to develop my own code. So, no problem if it's not currently available :)