The impact of socially desirable responding or faking on noncognitive assessments remains an issue of strong debate. One of the main reasons for the controversy is the lack of a statistical method to model such response sets. This article introduces a new way to model faking based on the assumption that faking occurs due to an interaction between person and situation. The technique combines a control group design with structural equation modeling and allows a separation of trait and faking variance. The model is introduced and tested in an example. The results confirm a causal influence of faking on means and covariance structure of a Big 5 questionnaire. Both effects can be reversed by the proposed model. Finally, a real-life criterion was implemented and predicted by both variance sources. In this example, it was the trait but not the faking variance that was predictive. Implications for research and practice are discussed. (Contains 2 figures and 3 tables.)