CAT with two-tier model (bifactor)

[janine.f...@gmail.com]

Hello Phil,

at the moment I try to run a CAT simulation with a bifactor-model and a two-tier bifactor model.
I have two questions regarding this try:
1) The bifactor model has one general factor and three specific factors. When I compared two cycles - one with defined weights for nuisance factors (https://groups.google.com/forum/#!searchin/mirt-package/nuisance%7Csort:date/mirt-package/ed_WoByHJzk/yM9FbdBfBQAJ) and one without weights  - the one without weights needs less items to achieve the same SE_theta.
Is this plausible ?

2) I tried to run a two-tier bifactor model (2 correlated general factors and three specific factors in each case). I defined min_SEM like

result_CAT <- mirtCAT(mo=A_bif, local_pattern=A, progress=TRUE,method = 'MAP',criteria="Drule",start_item = 'Trule',design = list(min_SEM=c(0.4,0.4,99,99,99,99,99,99)))

The problem is the first items ("items_answered") are in any case in the same ascending order like 1,2,3,4,5,6...

Quite sure I miss something important but am not able to solve this strange behaviour.

I would be very glad, if you could give me a hint.

Best 

Janine

[Phil Chalmers]

On Mon, Feb 26, 2018 at 8:14 PM, <janine.f...@gmail.com> wrote:

Hello Phil,

at the moment I try to run a CAT simulation with a bifactor-model and a two-tier bifactor model.
I have two questions regarding this try:
1) The bifactor model has one general factor and three specific factors. When I compared two cycles - one with defined weights for nuisance factors (https://groups.google.com/forum/#!searchin/mirt-package/nuisance%7Csort:date/mirt-package/ed_WoByHJzk/yM9FbdBfBQAJ) and one without weights  - the one without weights needs less items to achieve the same SE_theta.
Is this plausible ?

It depends on convergence criteria, but of course it could just happen this way from the selection of items given the corresponding theta estimates. 

 

2) I tried to run a two-tier bifactor model (2 correlated general factors and three specific factors in each case). I defined min_SEM like

result_CAT <- mirtCAT(mo=A_bif, local_pattern=A, progress=TRUE,method = 'MAP',criteria="Drule",start_item = 'Trule',design = list(min_SEM=c(0.4,0.4,99,99,99,99,99,99)))

The problem is the first items ("items_answered") are in any case in the same ascending order like 1,2,3,4,5,6...

Quite sure I miss something important but am not able to solve this strange behaviour.

You don't want Drule here, it doesn't take into account the structure of the latent traits, and the determinant of a sparse matrix is 0. Hence, mirtCAT just picks the next item in series, because all Drule values are equal to 0.

 

I would be very glad, if you could give me a hint.

Best 

Janine

--
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+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[janine.f...@gmail.com]

Hi Phil,

thanks for the answers!

Am Dienstag, 27. Februar 2018 16:52:01 UTC+1 schrieb Phil Chalmers:

On Mon, Feb 26, 2018 at 8:14 PM, <janine.f...@gmail.com> wrote:

Hello Phil,

at the moment I try to run a CAT simulation with a bifactor-model and a two-tier bifactor model.
I have two questions regarding this try:
1) The bifactor model has one general factor and three specific factors. When I compared two cycles - one with defined weights for nuisance factors (https://groups.google.com/forum/#!searchin/mirt-package/nuisance%7Csort:date/mirt-package/ed_WoByHJzk/yM9FbdBfBQAJ) and one without weights  - the one without weights needs less items to achieve the same SE_theta.
Is this plausible ?

It depends on convergence criteria, but of course it could just happen this way from the selection of items given the corresponding theta estimates. 

So this means more or less to try it out and just see what works better

 

2) I tried to run a two-tier bifactor model (2 correlated general factors and three specific factors in each case). I defined min_SEM like

result_CAT <- mirtCAT(mo=A_bif, local_pattern=A, progress=TRUE,method = 'MAP',criteria="Drule",start_item = 'Trule',design = list(min_SEM=c(0.4,0.4,99,99,99,99,99,99)))

The problem is the first items ("items_answered") are in any case in the same ascending order like 1,2,3,4,5,6...

Quite sure I miss something important but am not able to solve this strange behaviour.

You don't want Drule here, it doesn't take into account the structure of the latent traits, and the determinant of a sparse matrix is 0. Hence, mirtCAT just picks the next item in series, because all Drule values are equal to 0.

From your experience: Do you have any criteria (like Kullback-Leibler) in mind or is this also simulate and then choose what works best?

  

Best Janine

[Phil Chalmers]

On Tue, Feb 27, 2018 at 12:23 PM, <janine.f...@gmail.com> wrote:

Hi Phil,

thanks for the answers!

Am Dienstag, 27. Februar 2018 16:52:01 UTC+1 schrieb Phil Chalmers:

On Mon, Feb 26, 2018 at 8:14 PM, <janine.f...@gmail.com> wrote:

Hello Phil,

at the moment I try to run a CAT simulation with a bifactor-model and a two-tier bifactor model.
I have two questions regarding this try:
1) The bifactor model has one general factor and three specific factors. When I compared two cycles - one with defined weights for nuisance factors (https://groups.google.com/forum/#!searchin/mirt-package/nuisance%7Csort:date/mirt-package/ed_WoByHJzk/yM9FbdBfBQAJ) and one without weights  - the one without weights needs less items to achieve the same SE_theta.
Is this plausible ?

It depends on convergence criteria, but of course it could just happen this way from the selection of items given the corresponding theta estimates. 

So this means more or less to try it out and just see what works better

Basically, yes. If you knew something about the population then you might be able to come up with a better strategy "a la" Bayesian-esk approaches.

 

2) I tried to run a two-tier bifactor model (2 correlated general factors and three specific factors in each case). I defined min_SEM like

result_CAT <- mirtCAT(mo=A_bif, local_pattern=A, progress=TRUE,method = 'MAP',criteria="Drule",start_item = 'Trule',design = list(min_SEM=c(0.4,0.4,99,99,99,99,99,99)))

The problem is the first items ("items_answered") are in any case in the same ascending order like 1,2,3,4,5,6...

Quite sure I miss something important but am not able to solve this strange behaviour.

You don't want Drule here, it doesn't take into account the structure of the latent traits, and the determinant of a sparse matrix is 0. Hence, mirtCAT just picks the next item in series, because all Drule values are equal to 0.

From your experience: Do you have any criteria (like Kullback-Leibler) in mind or is this also simulate and then choose what works best?

Most likely just simulate. Though others in the literature might have better advice. 

Phil