Methodological Developments within the Quasi-Direct Format Demand Structure : the Multicountry Application for Passengers MAP-1

Abstract

In this paper, we perform four tasks. FIRSTLY, we summarise the seminal fixed-form formulations of coupled two-stage models initiated 30 years ago in the context of Northeast Corridor Transportation Project of the U. S. Department of Transportation and show how the use of Box-Cox transformations made it possible a decade later both to unify the various model formulations and to challenge the credibility of model results obtained under fixed-form component assumptions, notably in gravity-type models of levels and in classical linear Logit-type choice models. SECONDLY, we note that, as soon as either the Level or the Split component of two-stage models are considered separately and can be effected either with aggregate or with disaggregate data, the four possible combinations define a structure, the Quasi-Direct Format, with four matching classes of algorithms. THIRDLY, we summarise particular streams of improvement of three of these classes, all found in the TRIO program, Version 2.0, and concerned with the generalisation of so-called « root » models in different dimensions : in models of levels, ordinary least-squares procedures are generalised with respect to form of the fixed part and to heteroskedasticity and autocorrelation of the residuals ; in split models, the usual linear Logit model is generalised to non-linear forms that enrich the utility function and solve endemic parameter underidentification problems, and to the addition of choice captivity or modeler ignorance measures. We distinguish between background baseline pre-1996 developments and improvements under STEMM since 1996 : in models of levels, the algorithmic generalisations allow for tests of spatial autocorrelation of residuals with large samples ; in models of splits, numerous extensions of the Logit choice model are made possible with disaggregate data. FOURTHLY, we describe the first-ever multicountry application for passengers of the QDF format, obtained using the TRIO baseline algorithms : MAP-1 consists in a combination of disaggregate mode choice models and aggregate total demand models pertaining to three trip purposes. We also perform advanced tests of the algorithmic improvements effected under STEMM and demonstrate their usefulness on the business trip components of the MAP-1 multicountry application within the QDF structure. Key-words : travel demand, mode choice, Northeast Corridor, Germany, Europe, INSEE, International passenger Survey, Civil Aviation Authority, regression analysis, Box-Cox, Box-Tukey, Inverse Power Transformation, asymmetry of response, Logit, Dogit, captivity, ignorance, fidelity, heteroskedasticity, serial autocorrelation, directed autocorrelation, spatial autocorrelation, TRIO program, elasticities, probability points, parameter invariance, maximum likelihood, MAP-1 model. Resume Nous realisons quatre tâches dans cet article. EN PREMIER LIEU, nous resumons les formulations a forme fixe des modeles a deux etapes couplees elabores il y a une trentaine d’annees dans le cadre des etudes de la demande de transport dans le couloir Nord-Est des Etats-Unis sous l’egide du ministere americain des transports et nous rappelons qu’une dizaine d’annees plus tard il a ete possible, grâce aux transformations de Box et Cox, d’en unifier les formulations tout en mettant en doute leur credibilite dans la mesure ou leurs resultats avaient ete obtenus en supposant pour leurs composantes des formes fixes, notamment la forme gravitaire pour les modeles de niveaux et lineaire pour la composante de repartition de type Logit classique. Nous faisons remarquer EN SECOND LIEU que des qu’on distingue chaque composante de l’autre et que le modele de niveaux et le modele de repartition peuvent chacun etre construits avec des donnees agregees ou desagregees, la structure ou Format Quasi-Direct qui en resulte definit quatre combinaisons possibles auxquelles correspondent quatre classes d’algorithmes. Nous resumons EN TROISIEME LIEU, pour trois de ces classes, presentes dans le progiciel TRIO en Version 2.0, des directions d’amelioration axees sur la generalisation de « modeles-racines » dans certaines dimensions. Dans les modeles de niveaux, la generalisation des moindres carres ordinaires se fait par l’introduction de formes souples sur la partie dite fixe du modele et la prise en compte de l’heteroscedasticite et de l’autocorrelation des erreurs. Dans le modele Logit de choix, elle se fait par l’usage de formes non-lineaires qui permettent l’enrichissement des fonctions d’utilite et resolvent les problemes inherents d’identifification des parametres, ainsi que par l’ajout de parametres qui mesurent la captivite aux alternatives ou l’ignorance de l’analyste. Nous distinguons entre les ameliorations anterieures a 1996 et celles qui, depuis, sont imputables a STEMM: il s’agit,

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 4 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.