Program dependence is a fundamental concept to many software engineering tasks, yet the traditional dependence analysis struggles to cope with common modern development practices such as multi-lingual implementations and use of third-party libraries. While Observation-based Slicing (ORBS) solves these issues and produces an accurate slice, it has a scalability problem due to the need to build and execute the target program multiple times. We would like to propose a radical change of perspective: a useful dependence analysis needs to be scalable even if it approximates the dependency. Our goal is a scalable approximate program dependence analysis via estimating the likelihood of dependence. We claim that 1) using external information such as lexical analysis or a development history, 2) learning dependence model from partial observations, and 3) merging static, and observation-based approach would assist the proposition. We expect that our technique would introduce a new perspective of program dependence analysis into the likelihood of dependence. It would also broaden the capability of the dependence analysis towards large and complex software.