Tony LI posted on Thursday, June 11, 2009 - 7:03 am
I am wondering is there a good way to infer item location using IRT parameters.
I am fitting 2PL GRM (Mplus)to a 12 items scale and intend to use thresholds to derive the items' location on the latent trait. The scale has both positive and negative items (I have not reversed the response score for negative items when estimating).
The response options are 4 point Likert scale ranging from "Not like me at all" to "Exactly like me". I use the last threshold as approximation of item location, which seem to work ok for positive items.
However, I seem to have problem with the negative items. For example a negative item has thresholds of -0.066, 1.446, 2.296. Using the 3rd threshold 2.296 as item location does not make sense, as a negative item should be on the negative end of the trait continnum. Intuitively, I am thinking should I reverse the threshold to -2.296 or should I use -0.066 to approximate the item location. Or am i completely on the wrong track?
For the negative items I assume you get negative loadings. So a threshold of approx. 0 (-0.066), implying that most people are in the Not like me at all category, refers to the category Exactly like me if you had reversed the item before analysis and gotten a positive loading. Which is a complex way of saying that I think you should use -0.066 to be in line with what you do for the positive items.
See equation (27) in https://www.statmodel.com/download/MplusIRT.pdf The steps are the tau parameters and Beta is the regression parameter reported in Mplus as U on X or Eta by U. The Mplus output corresponds to that parameterization. Take a look also at User's Guide example 5.5.