Back from the IV EAM conference that was held in Postdam, near Berlin. The next one is planned in two years in Spain. In the mean time, I expect great publications coming up from some of the presenters.
There was a lot of interesting talks although five parallel sessions inevitably led to tedious alternative forced choice decisions (unless one’s willing to run from one room to another in less than 2 minutes, but this is not my case). I particularly appreciated the talk given by Keith F. Widaman, “Testing theoretical conjectures directly using constrained regression analysis.” The talk started with a review of Meehl’s paradox and elaborated on the usefulness of always testing against the null hypothesis of a null effect. He rather emphasized the difference between null hypothesis as seen in Physics (where we are interested in how observations differ from predictions) vs. those classically found in psychology (where we most of the times test the “no effect” H0, within the Fisher’s framework for testing hypothesis). The idea was then to set up a different effect under H0 and use regression analysis following the assumption of fixed effect (e.g., constant ratio between two regression coefficients). Nice work to follow on…
I also attended Andries van der Ark’s talk, and I am now waiting for the shipping of his book on Marginal models. The basic idea is that most if not all of the models that are currently headed under the IRT framework might be seen as marginal models. I look forward to read the entire book.
There was also interesting discussions around CAT in clinical setting. It has been successfully applied when investigating anxiety (Walter et al, 2007) and depression (Fliege et al, 2005). Interestingly, N. Smits constrasted CAT as used in a measurement perspective vs. prediction or diagnosis context (i.e., where one only wants to assign a given patient to one of two diagonstic groups).
Also, Paul De Boeck’s keynote was greatly appreciated; it focused on categories (i.e., categorical attributes) and dimensions (i.e., unidimensional scales) and helped to highlight the close proximity between these two conceptual vision of what may be classified and what may be ordered on a metric scale. The basic conclusion is that both can be nested together, or even exchangeable. Mixture item response models constitute a promising approach in psychometrics; but we need to look at the third and fourth moments to see the difference between a latent profile model (with p parameters) and a factor model (with p-1 parameters).
As a conclusion, I rest on my idea that latent variables might be thought of as tools constructed to answer specific questions rather than true constructs envisioned under a generating model.