Item re-scored so that all values are within a distance of 1

[chest...@gmail.com]

Dear researchers,

First, I would like to thank Phil for making such an amazing programme mirt to do various complex IRT modeling. It is very useful and it is one of the best programmes to conduct the analysis.

When I tried out the mirt programme, I get the following message:

Item re-scored so that all values are within a distance of 1

Also, the programme is running extremely slow.

I briefly check the data and found that the values do not have decimal places. I wonder if this is because in certain items (with Likert scale from 1 to 5), for example, some options have never been chosen? What are the other possible reasons behind this warning?

Thank you,
Chester Kam

[Phil Chalmers]

Hi Chester,

On Mon, Apr 11, 2016 at 11:01 PM, <chest...@gmail.com> wrote:

Dear researchers,

First, I would like to thank Phil for making such an amazing programme mirt to do various complex IRT modeling. It is very useful and it is one of the best programmes to conduct the analysis.

When I tried out the mirt programme, I get the following message:

Item re-scored so that all values are within a distance of 1

Also, the programme is running extremely slow.

I briefly check the data and found that the values do not have decimal places. I wonder if this is because in certain items (with Likert scale from 1 to 5), for example, some options have never been chosen?

That's exactly the problem. If some responses are obviously missing then all higher scores are changed so that they are equally spaced. Cheers.

Phil

 

 

What are the other possible reasons behind this warning?

Thank you,
Chester Kam

--
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.

[chest...@gmail.com]

Thanks for your reply. I am trying to get fit statistics with the M2 command. M2 requires complete dataset with no missing. Therefore, I tried the following command but failed:

> M2(modeling.irt,impute=10,QMC=TRUE)
Error in rep.int(rep.int(seq_len(nx), rep.int(rep.fac, nx)), orep) : 
  vector is too large
In addition: Warning messages:
1: High-dimensional models should use quasi-Monte Carlo integration. Pass QMC=TRUE 
2: High-dimensional models should use quasi-Monte Carlo integration. Pass QMC=TRUE 
3: High-dimensional models should use quasi-Monte Carlo integration. Pass QMC=TRUE

I also tried imputation but it was said that the information matrix was not positive definite:

> careful.scores <- mirt::fscores(modeling.irt, method='MAP', MI=100)
Error: Information matrix is not positive definite

Each of my items has 5 ordinal categories (1=Strongly Disagree; 2 = Disagree; 3= Neutral; 4 = Agree; 5 = Strongly Agree). Some items have very sparse response in the Strongly Disagree category (less than 5), even that my sample size is close to 500. The number of items is 55. I wonder if it helps to collapse categories into 3 categories (e.g., 1 = Disagree; 2 = Neutral; 3 = Agree)?

Thanks again!

Chester Kam

[Phil Chalmers]

Sounds like there is a lot going on in this model. The information matrix not being PD might suggest the model isn't identified, so that could be a large problem, and it looks like you might have a large amount of missing data (in which imputation isn't a good idea anyway). Collapsing responses would certainly help with stability, but even then you may have issues if too many parameters are being estimated. HTH.

Phil

[Ki Cole]

I am doing a study of polytomous IRT, specifically studying situations with low response rates. I am trying to better understand the "Item re-scored so that all values are within a distance of 1."

You respond, "If some responses are obviously missing then all higher scores are changed so that they are equally spaced." 

Can you provide more insight into this adjustment or where I am read about it further. I'm traced the mirt source code, but I'm not able to locate the technicalities of this change.

Thank you.

Ki Cole

[Phil Chalmers]

The idea is that if the observed item's unique responses are of the form:

c(1,2,4) == unique(item)

where for some reason it's clear that the 3rd category was never endorsed, then mirt will change all instances of 4 to 3

c(1,2,3) == unique(item)

Hence, all the unique values are 1 unit appart. The reason is somewhat technical, but because IRT models assume categorical relationships the values of the coding scheme should be relevant, and this just helps the internal arguments determine the ordering and spacing of the categories. For most IRT models, such as the graded response, nominal, etc, this translation is irrelevant. For other models where the scoring scheme is meaningful (such as in a partial credit model) then the fixed scoring coefficients would need to be overwritten at the time of estimation. So, unless you are using itemtype = 'gpcm', then I wouldn't worry.

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+unsubscribe@googlegroups.com.

[eaez]

Hi Phil,

I am fitting rsm, pcm, gpcm, and ggum models to a set of datasets, and received the warning that "item re-scored so that all values are within a distance of 1". Can you help how to go about avoiding this?

Thanks! 

Phil

To unsubscribe from this group and stop receiving emails from it, send an email to mirt-p...@googlegroups.com.

[Phil Chalmers]

This means that there were unexpected spaces between responses, such as noticing the unique observations for a given item are [1,3,4,5]. The rescoring would change the observed values to [1,2,3,4] instead, and use that. The warning appears because for some models, like the rsm, this can provide the incorrect rating constraints and should be forced instead (e.g., see the customK input to the technical list). For other itemtypes though, this really isn't a big issue, and forcing the customK input for categories that were not observed could even make the model unidentified and fail to converge properly. 

Phil

To unsubscribe from this group and stop receiving emails from it, send an email to mirt-package...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mirt-package/ef00484b-a033-4d1c-9a70-d4db51156e08%40googlegroups.com.