Testability is an important quality characteristic of software. A lack of testability contributes to a higher test and maintenance effort. Metrics can be used locate parts of a program which contribute to a lack of testability. In this paper we present a new approach to define metrics for software dependencies. We use this approach in the context of testability to identify test-critical dependencies, i.e. those dependencies within a system that are critical for test complexity. The results of four case studies show that 1) a small subset of the dependencies within a system has an exceedingly high impact on particular testability metrics, 2) conventional coupling metrics are not good predictors of these test-critical dependencies, 3) dependencies automatically identified to be test-critical are good indicators of design that needs improvement.