A Disaggregate Model of Car Ownership Based on the French National Transport Survey
R Grimal, French Technical Department of Studies on Roads, Transports and Highways, FR
A model of car ownership is developed from the National Transport Survey in trying to understand the determinants of mobility, and applied on the national census data. Need for temporal parameterization is derived from observing behavioural change.
This work is part of a series of studies trying to get a better understanding of long-distance mobility determiners today, from the last French National Transport Survey. After a first stage mostly dedicated to descriptive data analysis, modelling works are now beginning. Car still remains today the main transport mode and the first way of access to long-distance mobility, as many moves may not be realized so easily by other modes. Long-distance mobility thus remains highly dependent on household car ownership, defined here as the number of cars available to the household for personal trips. From this perspective, one may try to understand determiners of car ownership today and develop specific models.
Our paper will display the development process of a disaggregate probabilistic discrete choice model of car ownership at the household level, based on data from the National Transport Survey. Several model shapes (multinomial, nested and cross-nested logit), including socioeconomic variables, description of housing place and environment (size of urban area, central or peripheric location...), and indicators of access to mobility (number of driving licenses, number of seasonal transit tickets...) are specified, estimated and compared, leading to select the most desirable model shape, influential variables and likely values of the parameters for a given specification. Results show similar goodness-of-fit of a segmented multinomial logit model based on the number of driving licenses within household, and a cross-nested logit where the one-car alternative is correlated both with the no-car and the two-cars alternative. In both models, the number of driving licenses appears to be by far the most influential variable. Other indicators of access to mobility, description of housing place and environment, and family structure also exert a significant influence. On the contrary, socio-economic variables don't reveal so highly discriminant as the second-hand market remains affordable for low-income households.
Models are then used for simulation on various sub-samples of the survey. At this stage, the focus is to develop a comprehensive model, looking for the best possible trade-off between goodness-of-fit, economy of variables, and predictive abilities. The next stage is to test it in forecasting appliance on exogenous data to check its robustness. Detailed files of the continuous census data from 2007 may be used for this purpose, as they both contain the level of household car ownership and some common explanatory variables with the National Transport Survey. Nonetheless, explanatory variables available in the National Transport Survey but not in the national census will have to be removed, leading to assess the resulting loss in goodness-of-fit. The number of driving licenses, in particular, isn't filled in the census files. An alternative solution is to build a new model from the national transport survey to simulate the number of driving licenses, which may depend on very few variables (household structure, age, gender and level of qualification of household members) available in the census data. Though not absolutely necessary from a modelling point of view, these two stages are better depicting the natural logic of access to mobility, through a series of successive filters : access to driving license, access to car availability, final access to car use through defined priorities of access between members of the household.
A similar model will be developed on data from the previous national transport survey in 1994. Preliminary data analysis, comparison between both models, and appliance of the model based on the previous survey to current data will show significant behavioural change which can't be reduced to socioeconomic, housing or urbanistic determiners. On one hand, continuous progress in car ownership even for given values of explanatory variables, and despite decreasing household size, reveals the growing need for individual autonomy within the household, especially for constrained mobility. On the other hand, new behaviours against the tide appear among young people and inhabitants of big urban areas. These results invalidate the feasability of using descriptive static models of car ownership in the perspective of forecasting purposes, and imply including temporal parameterization based on prospective assumptions about changing behaviours.
Association for European Transport