This paper deals with the broad-band frequency analysis of complex systems, characterized by the presence of numerous structural scales (flexible parts) attached to the skeleton of the structure (stiff part) and for which numerous local displacements – which are very sensitive to uncertainties – are then coupled with the usual global displacements. Due to this overlap of several scales of displacements, there is an overlap of the low-, medium-, and high-frequency regimes (LF, MF, HF). Hence the introduction of a multilevel reduced-order model (ROM), whose vector basis gathers LF-, MF-, and HF-like families of displacements, for which the separation proceeds from a given filtering strategy. Integrating the nonparametric probabilistic approach of uncertainties, the obtained multilevel stochastic ROM allows for assigning a specific statistical dispersion to each scale. The stochastic ROM allows for (1) tackling the dimensionality induced by the local elastic modes and (2) taking into account the heterogeneous uncertainties associated with the frequency regimes.