Measuring change in a latent variable over time is often done using the same instrument at several time points. This can lead to dependence between responses across time points for the same person yielding within person correlations that are stronger than what can be attributed to the latent variable. Ignoring this can lead to biased estimates of changes in the latent variable. In this paper we propose a method for modeling local dependence in the longitudinal 2PL model. It is based on the concept of item splitting, and makes it possible to correctly estimate change in the latent variable.