The broad aim of this research is to contribute knowledge that enables improvements in how designers model decision alternatives at the systems level—i.e., how they model different system configurations and concepts. There are three principal complications: (1) design concepts and systems configurations are partially-defined solutions to a problem that correspond to a large set of possible design implementations, (2) each concept or configuration may operate on different physical principles, and (3) decisions typically involve tradeoffs between multiple competing objectives that can include "non-engineering" considerations such as production costs and profits. This research is an investigation of a data-driven approach to modeling partially-defined system alternatives that addresses these issues. The approach is based on compositional strategy in which designers model a system alternative using abstract models of its components. The component models are representations of the rational tradeoffs available to designers when implementing the components. Using these models, designers can predict key properties of the final implementation of each system alternative. A new construct, called a parameterized efficient set, is introduced as the decision-theoretic basis for generating the component-level tradeoff models. Appropriate efficiency criteria are defined for the cases of deterministic and uncertain data. It is shown that the model composition procedure is mathematically sound under reasonable assumptions for the case of deterministic data. This research also introduces an approach for describing the valid domain of a data-driven model based on the use of support-vector machines. Engineering examples include performing requirements allocation for a hydraulic log splitter and architecture selection for a hybrid vehicle.