Developing a high-fidelity stochastic multi-scale modelling framework that can capture the structure–property relationship in a computationally efficient manner is of interest to the structural engineers working with composite structures. This work revisits the problem of defining representative volume elements (RVEs) for performing a combined computational homogenisation-dehomogenisation procedure to estimate critical material allowables of unidirectional fibre reinforced polymer (FRP) composites. Particularly, the size and shape effects of 3D RVEs are investigated by performing Monte Carlo simulation of RVE arrangements to analyse the statistical convergence of their geometric features and the predicted structural properties in comparison with commonly used unit-cell models. For the microstructural features adopted in this work, it has been demonstrated that appropriately sized RVE models in two different aspect ratios could be considered as 3D statistical volume elements (SVEs) to efficiently capture the properties of interest. Attention has also been drawn to the implications of domain shapes and aspect ratios when small SVE models are used in the modelling framework for regulating computational power.