vcov() for exploratory models

[Gina Velázquez]

Hi! 

Thanks Phil for developing such an useful package, it makes possible really interesting research! I do not have yet much experience with the package, and I'm encountering difficulties to obtain the var-covariance matrix by means of vcov() for mirt objects that are exploratory. Depending on the model I want to estimate, I get either 1x1 NA-matrix output, or an error message where  "Error in solve.default(info) : \n  Lapack routine dgesv: system is exactly singular: U[1,1] = 0\n"

An example of model specification call:

mirt(data = dp[, item_index], model = 2, method = "MHRM", itemtype = "2PL", SE = TRUE, SE.type = "MHRM")

The size of the data object here is I=3128 and J=17. In all of the confirmatory mirt models I get estimated vcov() and it works normally, so I wonder is just not possible to use upon exploratory mirt objects?

Thanks in advance,

Regina

[Phil Chalmers]

HI Regina,

It's possible to estimate the acov for exploratory models, but it's generally not useful (since you need to rotate the model, the standard errors are only computed for the initial orientation....which is pretty useless). I'm not entirely sure why it's not positive definite though, especially if your confirmatory models are (perhaps the larger number of pars just makes it too unstable), but I wouldn't worry much about it. Cheers.

Phil

On Tue, Jun 7, 2016 at 4:50 AM, Gina Velázquez <gin...@gmail.com> wrote:

Hi! 

Thanks Phil for developing such an useful package, it makes possible really interesting research! I do not have yet much experience with the package, and I'm encountering difficulties to obtain the var-covariance matrix by means of vcoc() for mirt objects that are exploratory. Depending on the model I want to estimate, I get either 1x1 NA-matrix output, or an error message where  "Error in solve.default(info) : \n  Lapack routine dgesv: system is exactly singular: U[1,1] = 0\n"

An example of model specification call:

mirt(data = dp[, item_index], model = 2, method = "MHRM", itemtype = "2PL", SE = TRUE, SE.type = "MHRM")

The size of the data object here is I=3128 and J=17. In all of the confirmatory mirt models I get estimated vcov() works normally, so I wonder is just not possible to use upon exploratory mirt objects?

Thanks in advance,

Regina

--
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.

[Gina Velázquez]

Hi Phil,

thanks for your answer. I see it does not make much sense since it is based the SE of a certain rotation. The reason why I'm interested in these exploratory models is that my confirmatory ones (based on simple  structure derived from exploratory models) do not present good fit. I wanted to implement the delta method for obtaining b parameter and it's confidence interval, for what I need the vcov(). Is there any way I could compute the CI for a multidimensional b estimate without using an vcov() object?

Thanks!

Regina