The behavior of earthquake stress-drop magnitude scaling has been the topic of significant debate in the earthquake source community over the past two decades. Methodologies which have been adopted by a large number of source studies require corrections for source radiation pattern, path attenuation and site amplification that ultimately introduce large uncertainties for stress-drop estimates. In this study, we adopt a different strategy: we analyze directly the ground-motions (Y) and their dependencies with magnitude (M). We first use simple stochastic models (e.g. [1]) comprised of a [2, 3] source spectrum and various models of magnitude-dependent stress drop. We show that magnitude-dependent stress-drop and constant stress-drop models lead to different scaling of ground-motions (dlogY/dM) with frequency. Using the results of [4], we then analyze the magnitude dependency of NGA-West 2 ground-motions for source-site configurations where stress-drop is the key controlling factor of ground-motions (moderate distances and rock-sites). In addition, the use of a neural network method allows us to obtain fully record-driven evaluations of (dlogY/dM) with frequency both for simulated and observed records. The comparison between these observed and simulated (dlogY/dM) allows us to discuss the scaling of the stress-drop with magnitude. We do not observe strong differences of the magnitude scaling of ground-motions between mainshocks and aftershocks.