If the most parsimonious behavioral model between an observed behavior, Y, and some factors, X, can be defined as f(Y|X₁, X₂), then f_x1 will measure the impact in behavior of a change in factor X₁. Additionally, if f_x1x2 ≠ 0, then the impact in behavior of a change in factor X₁ is qualified, or moderated by X₂. If this is the case, X₂ is said to be a moderating variable and f_x1x2 is said to be the moderating effect. When Y is modeled via a logistic regression, the moderation effect will exist regardless of whether the index function of the logit specification includes a moderation term or not. Thus, including a moderation terms in the index function will help the researcher more precisely qualify the moderation effect between X₁ and X₂. The question that naturally arises is whether the researcher must include the moderation term or not. In this document, we provide the conditions in which moderation terms will naturally arise in a logistic regression and introduce some modeling guidelines. We do so by introducing a general framework that nests models with no moderation terms in three scenarios for the independent variables, commonly found in applied research.

Within the INMOP 3 research project, an attempt was made to solve a number of problems associated with the methodology of modelling travel in urban areas and the application of intermodal models. One of these is the ability to describe the behaviour of transport system users, when it comes to making decisions regarding the selection of means of transport and searching for relationships between travel describing factors and the decisions made in regard of means of transport choice.

The paper describes a probabilistic approach to the determination of modal split, and the application of a logistic regression model to determine the impact of variables describing individual and mass transport travels on the probability of selecting specific means of transport. Travels in local model of Warsaw city divided into 9 motivation groups were tested, for which ultimately 8 models were developed, out of which 7 were deemed very well fitted (obtained pseudoR² was well above 0.2).