In partitioned analysis, constituent models representing different scales or physics are routinely coupled to simulate complex physical systems. Such constituent models are invariably imperfect and thus, yield a degree of disagreement with reality, known as model form error. This error propagates through coupling interfaces and degrades the accuracy of the coupled system. To efficiently improve the coupled system, resources must be allocated to systematically improve the constituent models. This study proposes a tool and an associated metric that exploits the availability of experimental data to prioritize constituent models. This metric is useful in tracing the error of coupled systems to their origins and to quantify the contribution of constituent error to the overall error of coupled systems. The proposed metric is used to rank constituents based on (i) the relative model form error of the constituents, (ii) the sensitivity of the model form error of the coupled system to the model form error in the constituents, and (iii) the cost to improve their performance. The applicability of the proposed metric is demonstrated through a proof-of-concept structural example, by coupling individual frame elements to model a portal frame. Coupling and uncertainty inference of the inexact constituent models are achieved using optimization, where both separate-effect and integral-effect experiments are employed to train the model form error of the constituents and coupled system.