This study aims to make an objective comparative analysis between the relative significance of three crucial modelling aspects involved in the probabilistic analysis of transport networks. The first question to address is the extent to which the size of generated path sets can affect the prediction of the static flow in the path-based traffic assignment paradigm. The importance of this question arises from the fact that the need to generate a large quantity of paths may be perceived by analysts as a preventative reason as to the application of path-based stochastic traffic assignment (STA) models for large-scale networks. A simulated path generation algorithm, which allows the number of generated paths to be under modeller's control, is applied. Findings show that the size of the generated path sets does not substantially affect the flow prediction outcome in this case study. Further investigations with respect to the relative importance of STA model estimation (or equivalently, parameter calibration) and model specification (or equivalently, error term formulation) are also conducted. A paired combinatorial logit (PCL) assignment model with an origin–destination-specific-parameter, along with a heuristic method of model estimation (calibration), is proposed. The proposed model cannot only accommodate the correlation between path utilities, but also accounts for the fact that travelling between different origin–destination (O–D) pairs can correspond to different levels of stochasticity and choice randomness. Results suggest that the estimation of the stochastic user equilibrium (SUE) models can affect the outcome of the flow prediction far more meaningfully than the complexity of the choice model (i.e., model specification).