3
$\begingroup$

I have built an GPCM IRT model using survey data, but have found significant local dependence (in violation of the LID assumption) using residuals() in R. I've found information online for dealing with LD in Rasch models-- you can combine them into "super-items", which makes them polytomous.

Can I do this with polytomous data? What exactly does this mean (is it an item with no question and 10 response categories)?

Finally, how do you know which items to combine? Unfortunately, nearly all of my items are scoring high with more than one other item. I did use EFA (and MIRT) to see whether perhaps I was missing a dimension, but I can't seem to find anything significant.

$\endgroup$

2 Answers 2

2
$\begingroup$

From what you've described, my inclination is that you have item redundancy. Look at how highly correlated the flagged polygamous items are. If they are highly correlated, then you can consider just dropping one (or some) of the redundant items.

Additionally, if you wish to check if this is a reasonable approach, run the models with different items from the subset and then compare the ability estimates for each individual. If they are very highly correlated, then you probably have redundant items.

$\endgroup$
1
  • 1
    $\begingroup$ Good instincts and good call! I have some that can be seen as redundant. I (think?) can't drop them, though, because some people answered one, some answered the other, and some answered both. $\endgroup$
    – karen
    Mar 27, 2018 at 23:05
1
$\begingroup$

I agree with @Gregg H, dropping items is one approach. However, an issue with such an approach is that your scale's reliability would decrease (due to deleting items). Another approach could be to use a bifactor model (e.g., see Cai, Yang, & Hansen, 2011; and Reise, 2012 for more information) , which you could easily apply using the mirt package's bfactor() function. Such an approach would allow you to keep all the items displaying meaningful local dependence and simultaneously account for it by including specific factors (also referred to as group or nuance factors) that are orthogonal to your primary factor of interest.

References

Cai, L., Yang, J. S., & Hansen, M. (2011). Generalized full-information item bifactor analysis. Psychological methods, 16(3), 221.

Reise, S. P. (2012). The rediscovery of bifactor measurement models. Multivariate behavioral research, 47(5), 667-696.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.