IRT Model for number of attempts before correct response

Question

I have a test where respondents are able to attempt each item as many times as required until they get it correct. I am looking to fit an IRT model to data from this test in order to look at the relative difficulties of different items, and how this maps onto person ability.

One possibility would be to think about the data in a binary sense of whether the person got the item correct in the first attempt. This would fit neatly onto the sort of 1PL/Rasch model that I'm well familiar with. However, it seems reasonable to assume that the number of attempts that the person takes in order to correctly respond is also related to the difficulty of the item (i.e., an item that generally takes 5 attempts to complete may be more difficult than one that generally takes 4).

Does anybody know of any extension to the Rasch model, or any other IRT model, that would allow me to estimate difficulty and ability parameters for a test of this sort?

Answer 1

Fitting a generalised partial credit/graded response model would work well here if you are sure of this phenomenon (if not, maybe consider a nominal response model for items you're unsure of). The idea is to represent the number of attempts it takes the person to answer as the 'observed item score', and use a monotonic IRT model to predict these category responses. So, if they answered correctly on the first try, score them as 1, or if it took 3 attempts, score as 3.

The only issue here is that the slopes associated with these items will be negative (because lower scores are associated with higher ability), so just flip them in your head to interpret the strength of the item. Otherwise, everything else in your IRT model should be the same.