The Rasch Model Still Does Not Fit

Certain criticisms of the use of the Rasch model have been queried in a paper by Bryce. We argue that these criticisms not only remain valid, but are reinforced by recent research on latent trait models. In particular, Bryce fails to present a clear conceptual argument and adopts a very inefficient procedure for testing empirically the fit of the Rasch model. A paper by Bryce (1981) takes issue with some of the reasoning in earlier papers (Goldstein & Blinkhorn, 1977, Goldstein, 1979) where it was argued that the Rasch model was generally unsuitable for use in educational assessment. He also presents empirical evidence in justification of the use of the Rasch model for test construction and analysis. While we believe there is a certain amount of common ground between us, we also think that Bryce has partly misunderstood the position we have adopted, and ignored some serious weaknesses of the usual type of empirical analysis which he uses in his paper. We first deal with the conceptual framework. The Rasch Model Assumptions There seems to be no disagreement between us that the Rasch model is unsuitable with respect to large-scale item banking. This is of course the use of Rasch with which we were most concerned, and it has been pointed out (Goldstein, 1979) that the Rasch model ought properly to be regarded as a special kind of factor analysis model with possibly an important exploratory role. Again, we think there is no disagreement between us on the question of the need to test the assumptions of the Rasch (or indeed any other) model empirically. Beyond this, however, we seem to part company and we now take up the points made by Bryce. Bryce accuses us of 'axiomatically' stating that the complexity of data must be greater than that allowed for by the Rasch model while also allowing that the validity of Rasch is open to empirical verification. What we actually suggested was