Towards a Comprehensive Framework for Panel Data Analysis
S Hess, ITS University of Leeds, UK; A Daly, ITS University of Leeds/RAND Europe, UK
This paper presents a unified framework for dealing with the many phenomena that possibly play a role in panel data
Data sets containing multiple choice responses for each individual are now used extensively, both in academic and practical contexts for understanding and predicting travellers? choices. The potential impacts of this repeated choice nature of panel data have been discussed in numerous papers, where authors have discussed the implications on actual behaviour, notably in terms of fatigue and boredom, and experience and habit formation. Additionally, the implications for modelling have been discussed in several papers, looking at the likely impact on the error structure of the models as well as the possible existence of complex patterns of taste heterogeneity and scale heterogeneity in such data. But to date, there is little or no consensus on which of the phenomena plays the most important role, and also what the scope is for confounding between the individual effects.
This paper starts by presenting an extensive overview and discussion of the likely impacts of the repeated choice nature of panel data on survey responses as well as the implications for model estimation. We then develop a modelling framework that jointly takes into account these phenomena. Importantly, our model framework aims to deal with all effects at the model estimation stage, hence avoiding as far as possible a posteriori correction approaches such as bootstrap or jack-knife. One aim of this new framework is to highlight the significant risk of confounding between different phenomena. The output of this work will be a highly flexible mathematical structure that has the capability to account for all known potential effects of the panel nature of the data. The general model will be constructed in such a fashion as to be able to reduce to simpler structures as any non-applicable processes drop out one by one. As an example, the model will be able to account for variations in tastes across choice situations for a given respondent. If no such variations are identified from the data, the model will reduce to a structure that assumes that tastes stay constant across choice situations for the same respondent. Efforts will be made to account for as many effects as possible in a deterministic manner, whereas, at present, a large number of effects are simply absorbed into the error term of the models.
A new SC dataset is currently being collected for the purpose of this paper, where particular emphasis is placed on collecting information on possible effects of the repeated choice nature of the data. While a single data set cannot be decisive in indicating the relative importance of behavioural effects, the design of this data set will allow several different aspects of the general modelling framework to be illustrated and tested. Of course, indications will also be available of the behavioural results obtained from the specific data collected.
Association for European Transport