Statistical modelling of discrete choices in discrete time with panel data.

Despite the major advances during the last ten years in procedures for understanding and predicting individual choice behaviour, with rare exception the emphasis has been strictly cross-sectional. Available non-cross sectional studies are typically characterised by either highly aggregate data series (e.g. quarterly nationwide economic and demographic time series, annual automobile registration series) or repeated single cross-section data on the same closed population but with no planned comparison at the time of the initial survey (with the consequent lack of control over sample specification; question range, wording and order; and conduct). Although these studies will continue to have an important role, especially at the level of univariate descriptive representation (e.g. general changes in trip frequency rates) and in long-range forecasting (e.g. stock of motor vehicles in 2010), such data is not sufficiently rich intertemporally to be able to distinguish the many sources of influence on individual (and hence market) choice behaviour. Improvements in predictive capability now require us to identify the causes (or at least sequences) of change. For example, to be able to separate the effects of persistent interindividual differences from real intertemporal relationships, a panel can provide the desired statistical leverage to sort out the causal patterns. The major focus of this paper is on identification of sources of heterogeneity, nonstationarity and intertemporal dependence, and ways of incorporating these sources of behavioural variation into a computationally tractable model specification. Although this approach will in general require us to distinguish the nature of time (continuous, discrete), the nature of choice (continuous, discrete) and the nature of intertemporal linkage (pooling, structural) this paper is specialised to discrete choice in discrete time. The other specifications are detailed in a longer paper (available on request).(a) for the covering abstract of the conference see IRRD 290118.