Implications of Task Complexity for Discrete Choice Analysis: a Comparative Study of Nested Iogit and PRETREE
WINNER OF The Neil Mansfield Award
BATLEY R, ITS, University of Leeds, UK
Nested Iogit [NL] (McFadden, 1978) is among the most popular models in discrete choice analysis. NL can be derived from the theory of random utility, which assumes utility maximisation. NL is a compensatory choice model. That is, in terms of utility, a ch
Nested Iogit [NL] (McFadden, 1978) is among the most popular models in discrete choice analysis. NL can be derived from the theory of random utility, which assumes utility maximisation. NL is a compensatory choice model. That is, in terms of utility, a change in one attribute can be compensated by a change in another.
It is argued that, in some contexts, the complexity of the discrete choice task may preclude utility maximisation. 'When faced with many alternatives (e.g., job offers, houses, cars), people appear to eliminate various subsets of alternatives sequentially according to some hierarchical structure, rather than scanning all of the alternatives in an exhaustive manner...Indeed, these considerations have led several theorists (notably Simon, 1957) to modify the classical criterion of maximization and to view the choice process as a search for an acceptable alternative that satisfies certain criteria' (Tversky and Sattath, 1979 pp542-543). Choice task complexity appears to be particularly relevant in the context of the household demand for alternative-fuel vehicles (AFVs). 'A complication for personal vehicle buyers has been continuous diversification of the vehicle market; buyers may choose from an increasing array of manufacturers, models, and styles. AFVs will multiply this diversity, making auto choice even more complex' (Turrentine and Sperling, 1991 p211 ).
Tversky and Sattath (1979) addressed the challenge of task complexity by proposing the PRETREE model. PRETREE is a hierarchical specification of the elimination by aspects [EBA] model (Tversky, 1972a, 1972b). Unlike MNL and NL, EBA and PRETREE are not derived from the theory of random utility, and do not assume utility maximisation. Rather EBA and PRETREE employ an attribute-based sequential elimination process, achieving a satisfactory solution with a modest amount of computation.
Whereas NL has been applied in a vast number of empirical studies, the same cannot be said of PRETREE. The paper addresses this issue, summarising a comparative empirical study of the two models, set in the context of the household demand for AFVs (Batley, 1999; 2000). A series of synthetic NL and PRETREE data sets were generated by simulation. The study consisted of four principal analyses. In Analysis 1, a NL model was estimated for each NL data set. In Analysis 2, a PRETREE model was estimated for each PRETREE data set. Analyses 3 and 4 considered the implications of estimating the 'wrong' model. That is, Analysis 3 estimated a PRETREE model for each NL data set, and Analysis 4 estimated a NL model for each PRETREE data set.
Association for European Transport