The Use of Pseudo Panel Data for Forecasting Car Ownership

While car ownership forecasting has always been a lively area of research, traditionally it was dominated by static models. To utilize the rich and readily available repeated cross sectional data sources and avoid the need for scarce and expensive panel data, this study adopts pseudo panel methods. A pseudo panel dataset is constructed using the Family Expenditure Survey between 1982 and 2000 and a range of econometric models are estimated. The methodological issues associated with the properties of various pseudo panel estimators are also discussed. For linear pseudo panel models, the methodological issues include: the relationship between the pseudo panel estimator and instrumental variable estimator based on individual survey data; the problem of measurement errors (and when they can be ignored) and the consistent estimation of dynamic pseudo panel parameters under different asymptotics. Static and dynamic models of car ownership are estimated and a systematic specification search is carried out to determine the model with best fit. The robustness of the estimator is investigated using parametric bootstrap techniques. As an individual household’s car ownership choice is discrete, limiting the model to linear form is obvious insufficient. This study attempts to combine the pseudo panel approach with discrete choice model, which has the distinctive advantages of allowing both dynamics and saturation but without the need for expensive genuine panel data. This does not seem to have been done before. Under the framework of random utility model (RUM), it is shown that the utility function of the pseudo panel model is a direct transformation from that of cross-sectional model and both share similar probability model albeit with different scale. This study also explores the various forms of true state dependence in the dynamic models and tackles the difficult econometric issues caused by the inclusion of lagged dependent variables. The pseudo panel random utility model is then applied to car ownership modeling, which is subsequently extended to take saturation into account. The model with the best fit has a Dogit structure, which is consistent with the RUM theory and is able to estimate the level of saturation and test its statistical significance. Both linear and discrete choice models are applied to generate forecasts of car ownership in Great Britain to year 2021. While the forecasts based on discrete choice models closely match the observed car stock between 2001 and 2006, those based on linear models appear to be too high. Furthermore, the results from nonlinear models are comparable to the findings in other authoritative studies, while the long term forecasts from linear models are significantly higher. These results highlight the importance of saturation, and hence the choice of model functional form, in car ownership forecasts. In conclusion, we make some comments about the usefulness of pseudo panel models.