Abstract

This session will demonstrate how to use links to downloads that enable various psychometric analyses to be carried out in Stata. Details on these the links appear in The Psychometrics Centre website. One of these is Mokken scale analysis, a non-parametric hierarchical scaling method similar to Guttman scaling. Both techniques assume the existence of an underlying latent (unobservable) attribute, which is represented by a set of items that are ordered in terms of difficulty and related to a latent attribute. Both imply that any individual who agrees with a particular item will also agree with all the items ranked lower in difficulty. The main difference between Guttman scaling and Mokken scaling is that Mokken scaling has a probabilistic nature whilst Guttman scaling does not. In Mokken scale analysis reproducibility is measured by Loevinger’s coefficient Hi for each item and H for the entire scale. The calculation of Hi and H depend on comparing the probability of errors in ranking to the probability of such a ranking occurring if the items are unrelated. Hi and H will take values between 0 and 1. A commonly used rule of thumb is that a strong scale is one in which all Hi and H exceed 0.5 in value. Items can be omitted from a scale items if they have low value of Hi, hence it is a straightforward non-parametric approach to item analysis within test construction.