An important advantage of stated preference (SP) discrete choice experiments is that several responses can be collected from each individual. This reduces substantially the cost of data collection and allows for more advanced experimental designs. However, the collection of multiple responses means that each respondent's basic preferences apply to the series of responses that he or she has given: those responses are therefore interdependent (serially correlated). Naive analysis methods that assume the independence of observations are therefore, in principle, invalid. A number of methods can be used to correct for the interdependence of SP observations, including methods that capture these effects through sub-sampling procedure and models with specifications that explicitly estimate terms that capture the serial correlation among observations. Another possible approach can beto use an imploded sample where all responses from an individual are regarded as a single observation with multiple choices and each set of possible choice combinations is regarded as an alternative in the estimation. Each approach has benefits and drawbacks, both theoretical and practical. However, to date little research has been published comparing the results from these different procedures. This study therefore presents an empirical analysis to evaluate the performance of each of the methods, using a real-world transport application to compare and contrast the benefits and disbenefits of each approach. Comparisons are made between simple models with generic coefficients only and models with explicit consideration of observedtaste variation. The effects of serial correlations are then accounted for separately through each of the competing methods and incorporating individual-specific error terms in the model structures (restricted random coefficient models). The latter tests have been undertaken using several commercially available software packages. For model evaluation the estimated coefficients of the corrected random coefficient models are compared with the coefficients of the base (naive) models. Ease of use and estimation timeare also discussed. For the covering abstract see ITRD E145999
Abstract