The cost of assuring test quality significantly increases when dealing with complex systems with tightly integrated AMS-RF building blocks. Machine learning-based test may be a promising solution to this issue. These tests rely on regression models trained to replace costly performance mea- surements by simpler test signatures. However, these regression models are targeted only at parametric performance variations in defect-free circuits. The presence of spot defects may be undetected by these tests and lead to test quality degradation and reliability issues. In this work we propose a methodology based on causal discovery algorithms to screen out these spot defects.