Dear colleagues working with item analysis and test construction!
With my colleagues, we came up with some pretty awsome new formulae related to item-test and item-rest correlation. One of the results is that the traditional rule of thumb to reject items with Rit < 0.20 is too low. For a item with medium difficulty, the zero Rit is 1/SQRT(number of items), i.e., with 9 items in a (sub)test, Rit = 0.30 refers to rho = 0. Then, we should use the boundary 0.30 + 0.20 = 0.50 to select items with a correlation higher than 0. I invite you to read the preprint at Preprint Some hidden characteristics of item-total correlation and He...
.All possible feedback is taken positively.
Yours,
Jari M
Turku Research Institute for Learning Analytics (TRILA)