Abstract A comprehensive autoregressive (AR) and linear multistage autoregressive moving average (LMS-ARMA) framework for effective multichannel identification of structures under unobservable excitation is introduced and assessed via the paradigm of aircraft skeleton structure identification. This framework utilises fully parametrised vector AR and ARMA models estimated via linear regression (LR) and a newly developed linear multistage (LMS) method, respectively, statistical model order selection, and structural mode distinction via a dispersion analysis methodology. The framework's effectiveness and limitations are examined via the identification of an experimental and a simulated aircraft skeleton structure, as well as comparisons with non-parametric Welch autospectral density estimation. The results of the study indicate the framework's effectiveness in selecting the necessary model orders, as well as in distinguishing and accurately estimating most structural modes, including closely spaced and ‘local’ modes. The AR an LMS-ARMA methods achieve generally similar accuracy, but the latter leads to lower order models, fewer spurious modes, and less ambiguous ‘stabilisation’ diagrams.