The Sample Strategy Of A Test Information Function In Computerized Test Design

When the Rasch model is the IRT model, typically about five specification points suffice to prevent that the information of the constructed test has local minima between specification points considerably lower than the adjacent specified minima. For a typical test length this implies that the asymptotic property of the information function as the inverse of the variance-function of the maximum likelihood estimator is an adequate interpretation and justification for the specification of a minimum information function. As a consequence of the linear approach to optimization one has to restrict the specification of measurement
accuracy to a few important ability classes. Therefore, it seems natural to choose the specification points at values of the latent variable that can be considered to represent these classes.

However, when the IRT model allows different discrimination parameters or indices, like the OPLM model, local minima of the information function within a class may be very much less than the information at the specification point, and, moreover, may be responsible for too large deviations from the asymptotic
property to justify the above procedure. Therefore, a more general interpretation of the information function and an associated procedure for its specification in test construction is proposed that proves also adequate for
OPLM or the Birnbaum models. The interpretation clearly shows that the information has to be specified at edges that separate ability classes, not at abilities that represent classes.