The car market is an important sector of most modern economies and both the demand for ownership and the demand for new cars play an important role in economic decisions. The health of the car industry in general depends on consumers’ demand for new cars. Demand forecasting is one of the most important tools car manufacturers use in their financial planning and decision making about expansions and contractions of plant capacity. For various government and public bodies, understanding and forecasting demand for car ownership are equally important. As the most important user of petroleum fuel, the car market has a strong influence on non-replaceable energy.
Projections of future fuel consumption, and the impacts on fuel consumption of various forms of government intervention, are routinely based on forecasts for car ownership demand. Furthermore, understanding the factors driving demand for cars is important in addressing a range of environmental issues including local air pollution and climate change. Since car emissions are a large component of pollution, air quality standards and policies are largely based upon projected car ownership and use. Finally, accurate car demand models are also an aid to planners who must anticipate infrastructure needs, address concern of congestion and provide public transport services. Government agencies and local passenger transport authorities utilize projections of car ownership levels as a key input to obtain accurate projections of infrastructure needs and public transport patronage.
Car ownership forecasting plays a central role in the planning and decision making of numerous public agencies and private organizations. Given the important role of car demand forecasts in a wide variety of settings, it is not surprising that it has been a lively area of research and numerous models have been constructed to forecast car demand. It is important to recognize that the choices of model structure and functional form are heavily influenced by the objective and context of the study, and there is no single model that would offer best performance in all situations. For example, for short term forecast, a simple time series model based on aggregate data might perform better than the more complex disaggregate model. Also, while the consideration of saturation might not add much value for car forecasting in developing countries, it might be
highly significant in mature markets. It is easier to understand these points by looking at the car market in Great Britain as an example. Between 1950 and 2005, the total number of cars in the stock increased from 1.98 million to 26.21 million, implying an average growth rate of 4.7% per annum; for the same period, the real Gross National Income (GNI) increased from £243 billion to £987 billion (1995 prices), implying an average growth rate of 2.5% per annum (DfT, 2006c; ONS, 2007). If one uses certain time series models (for example, a simple Error Correction Model) for long term forecasts, it would substantially over-estimate the car stock in distant future. This is because the past growth trend will be inevitably curtailed by the approach of saturation:
in 1951, only 14% of households had regular access to at least one car, while this proportion increased to 75% in 2004 (DfT, 2006c).
In the current study, the car demand forecasting model is developed within the context of British car market. In the UK, the Department for Transport has commissioned a number of “official” forecasting models over the past few decades, which include those developed by the Transport and Road Research Laboratory, the Regional Highway Traffic Model (RHTM), the National Road Traffic Forecasts (NRTF) and the National Transport Model (NTM). Besides the Department for Transport, there are industrial organisation such as SMMT (The Society of Motor Manufacturers and Traders) and other commercial and academic organisations, which are also involved in this area of research. However, some of the research remains “in house”, i.e., the details of the model are not publicly available. Some of the studies available in the academic journals used methodology and data different to the NRTF/NTM, yet each study has its own limitations. Various types of car demand forecasting models developed in Great Britain and worldwide will be reviewed in Chapter 2.
The literature review reveals that the static approach dominates car ownership forecasting in Britain. The motivation for this thesis is that the inclusion of dynamics will yield fruitful results and lead to more accurate forecasts. Traditionally, empirical models of individual travel choice behaviour have been built on the assumption of equilibrium and suffered from a lack of dynamics. In the past two decades, the importance of dynamics in transport is gradually gaining recognition. In various areas of transportation research, issues such as the temporal dependence of choices, the role
of habit, imperfect information regarding alternatives and prices, costs of adjustment and transaction costs, have been empirically assessed.
Nevertheless, the use of dynamic approach in car demand forecasting is still limited due to heavy data requirements. Due to data constraint, there have been relatively few forecasting models that use the dynamic approach except those using aggregate time series methods. It is possible to forecast car demand using panel data models. However, there is only one panel survey in Britain containing limited transport related information: the British Household Panel Survey (BHPS), which is inadequate for the purpose of our study. Furthermore, due to the attrition problem, the size and representativeness of the samples decline over time, rendering the panel data inferior to other national cross-sectional data. For example, less than half of the respondents in Wave 1 of BHPS remained in the sample in Wave 13, and various population groups such as the old, the young, the unemployed, those with low income, etc. became significantly under-represented (ISER, 2006).
One approach to circumvent the need for panel data is to construct pseudo panels from the cross sectional data. The pseudo-panel approach is a relatively new econometric approach to estimate dynamic demand models. A pseudo-panel is an artificial panel based on (cohort) averages of repeated cross-sections. The cohorts are defined based on time-invariant characteristics of the households and extra restrictions should be imposed on pseudo-panel data before one can treat it as genuine panel data. Using the cohort data over a number of periods, one could distinguish long run and short run effects while allowing for heterogeneity between the cohorts. In this way, one is able to overcome the deficiencies in both the static models and aggregate time series.
The application of pseudo panel car ownership model raises many interesting questions.
For example, how do we define the cohort so the econometric model is identified and the measurement errors in variables can be minimized (ignored)? There is a question about the treatment of the dependent variable. Whereas in the original data car ownership is a discrete variable (zero, one, two, ..) in pseudo panel data it is a continuous variable, e.g. average number of cars per household or the proportions of households that own cars. There is a question about the treatment of transformations of the variables. Should one use the average of the transformed micro data or transform
the pseudo panel cohort data? In many cases, e.g. logarithmic transformations, the average of the transformed data is not defined, since the micro data contain zeros1. Is it possible to apply the microeconomic theory of utility maximization for individual decision makers and combine the pseudo panel model with the random utility model?
What are the pros and cons of discrete choice (nonlinear) pseudo panel model and what’s its relationship to standard random utility models? How can the nonlinear pseudo panel model be consistently estimated? And finally, what are the empirical appeals of the pseudo panel models and how well do they perform in car ownership forecasting?
To facilitate readers’ understanding, we list the most common notations that are used throughout the thesis in Table 1-1. The following subscripts are consistently used: i denotes the individual household in the micro survey; c denotes the associated cohort of household i; and t denotes years.
Table 1-1 Common Notations
Notation Description
Act Average number of cars (automobiles) per household in cohort c in year t P1+ Probability of household owning at least one car
P2+|1+ Probability of household owning two or more cars conditional on owning at least one car
nct Number of sample observations in cohort c in year t
Nc Total population of cohort c (assumed to be constant in the theoretical model, i.e. no birth or death)
C Total number of cohorts
Scalar: coefficient for the lagged dependent variable
✁ K x 1 vector of coefficients for exogenous explanatory variables
✂
Unobserved cohort heterogeneity (fixed effect or random effect)
In the empirical work, the dependent variable is Act for all linear models. It is slightly more complicated for discrete choice models. We observed the proportion (not probability) of households owning at least one car in cohort c in year t, which is noted as rct1+. Among the car owning households in cohort c, we also observed the proportion of those owning two or more cars, which is noted as rct2+1|+. They are the dependent
1 There are also theoretical considerations on the treatment of variable transformations. Again, it remains a question whether the method proposed in the standard linear pseudo panel econometric literature
variables in two separate discrete choice pseudo panel models. The first order condition of the maximum likelihood function implies that the predicted probability (P1+ and P2+|1+) equals the proportions (rct1+ and rct2+1|+) in the sample under certain conditions (e.g. probability model is a multinomial logit model with alternative specific constant).
The thesis is organized as follows. Chapter 2 is the literature review of car ownership models. Chapter 3 describes the data used in the thesis including the construction of the pseudo panel dataset. Chapter 4 discusses the linear static fixed effect models, investigating the relationship between the pseudo panel estimator and the instrumental variable estimator based on individual survey data as well as the measurement error problems (and when they can be ignored). Chapter 5 discusses the consistent estimation of linear dynamic pseudo panel model under different asymptotics and the rank conditions for identification. For empirical models of car ownership, systematic specification search is carried out to investigate various issues such as appropriate explanatory variables, functional forms, problems of heteroskedasticity and autocorrelation, fixed or random effects and presence of heterogeneity.
Chapter 6 extends the pseudo panel approach to discrete choice model. The pros and cons of nonlinear pseudo panel model are discussed and a pseudo panel model that is consistent with random utility theory is presented. Chapter 7 investigates dynamic discrete choice model of pseudo panel. Models with different forms of (true) state dependence are compared and consistent estimation methods are proposed for the preferred first order Markov model. For the car ownership model, saturation is an important concept so the theoretical model has been extended and a pseudo panel Dogit model is presented. Empirical models of households with at least one car and those with two or more cars conditional on owning the first car are estimated separately.
Chapter 8 uses both the linear and nonlinear econometric models to generate car ownership forecasts for Great Britain between 2001 and 2021. The forecasting results are compared to the observed car stock between 2001 and 2006 as well as the forecasts in other authoritative studies. A number of scenario tests are carried out to examine the sensitivity of the forecasting models. Chapter 9 is a brief conclusion, where the usefulness of pseudo panel models is also considered.