The majority of current empirical educational research is cross-sectional in nature. Studies of this kind do not allow evidence regarding the development of students’ achievement. In order to gain empirical information on educational processes longitudinal studies assessing the same students at different points of time are needed. From the methodological point of view however the implementation and the analysis longitudinal studies implicate numerous challenges.

Modelling repeated measurements with IRT scaling allows avoiding massive problems the classical test theory cannot prevent. Bereitner (1963) and Lord (1963) postulated that (1) the reliability of the difference between the tests show negative relation to the correlation between tests. That means the lower the correlation between the tests – the higher the reliability of difference. But high correlation between tests of different time points is desirable. (2) Different qualifications of the persons tested cannot be taken into account, since person and items difficulty are measured on one scale. (3) The shared error variance of the scores of the first measure point and the differences cause negative correlation between first measure point scores and differences, even when no change appears.

Especially for modelling test scores and data of competence testing IRT offers a sum of elegant methods, that eliminate or at least reduce the above mentioned problems. Currently multiplicities of models exist to measure change, which are realized in different applications:

(1) The Linear Logistic Test Model (LLTM or linear logistic RM) introduced by Fischer (1983, 1995a, b) and Spada & McGaw (1985) has the basic assumptions that the person parameter is considered constant over time and change in the overall item difficulty is equivalent to global change in latent person ability. Therefore change is homogeneous across persons which lead to the condition of one-dimensionality. The position of person v on the one latent trait underlies his or her response to all items i at all occasions to another. This model can be realized with the applications which are called virtual items or virtual persons. 

(2) The Linear Logistic Testmodel with Relaxed Assumptions (LLRA) by Fischer (1983, 1995c) and Andersen (1985) follows the assumptions that the change is person-specific and the item difficulty remains constant over time. Item form a one-dimensional scale at each time point which leads to a multidimensional latent space for each latent-trait continuum. Also item parameters are constant over time points. Virtual items or one test-treaty are applications to implement this model.

(3) The Loglinear conditional Rasch Model by Cressie & Holland (1983) and Kelderman (1983) models homogeneous change, with no person change considered. It is expressed without latent-person parameter, because the total score represents the person parameter. The expected probabilities of the response vectors are reparameterized as linear combinations of item parameters.

(4) Multidimensional Rasch Model for Learning and Change (MRMLC) by Embretson (1991) models change as a latent dimension takes person qualification into account.

This symposium gathers an international panel of speakers. The altogether six papers – organized in two sessions – address specific methodological challenges in modelling longitudinal data in educational research.