PUB - Publikationen an der Universität Bielefeld

Discrete choice models (DCMs) are ubiquitous in various disciplines, particularly in transportation planning, where they inform high-stakes infrastructure investments and policy decisions. However, the subjective nature of utility function specification has the potential to introduce variability in model outcomes, thereby undermining forecast reliability and potentially leading to discrepancies between data-driven recommendations and stakeholder decisions. This dissertation aims to address these challenges by means of three interconnected research papers, which collectively contribute to the development of more robust, transparent, and reliable DCMs.

The first paper introduces score plots, a diagnostic visualisation tool for models estimated using composite marginal likelihood (CML) methods, to detect model misspecifications and enhance model fit.

The second paper demonstrates the use of pairwise CML estimation for gradient-based Lagrange multiplier-type tests, enabling the comparison of gradients between different CML contributions and facilitating the testing of pooling over time and the identification of neglected temporal correlation. The third paper explores the effective use of power weights in CML estimation, particularly in unbalanced panel settings and in the presence of unaccounted autoregressive error structures, to reduce the variance and asymptotic bias of the estimator.

Collectively, these papers furnish a comprehensive methodological framework for enhancing the reliability and transparency of DCMs. By enhancing model specification, testing, and estimation, this research reinforces the credibility of data-driven recommendations and contributes to more informed decision-making. The dissertation highlights the potential of CML-based methods to navigate the intricacies of model specification, reduce subjectivity, and improve the overall quality of DCM-based forecasts and recommendations.