There are various models in software engineering that are used to predict quality-related aspects of the process or artefacts. The use of these models involves elaborate data collection in order to estimate the input parameters. Hence, an interesting question is which of these input factors are most important. More specifically, which factors need to be estimated best and which might be removed from the model? This paper describes an approach based on global sensitivity analysis to answer these questions and shows its applicability in a case study on the COCOMO application at NASA.