On the (im-)possibility of deriving transport policy implications from hybrid choice models

Although the notion that attitudes and perceptionsplay an important role in explaining travel choice behavior has been around fordecades, the incorporation of these latent factors in discrete choice models isfairly recent. Especially during the last few years, an increasing number ofresearchers have begun to develop and test such hybrid choice models, motivatedby the idea that the integration of latent attitude- and perception-relatedvariables in choice models enhances their behavioral realism and may ultimatelylead to more tailored and better informed travel demand policies. However, in this paper we argue that when it comes toderiving policy implications from these hybrid choice models (from here on:HCMs), their added value compared to that of conventional choice models israther limited, and that many recent papers in fact have presentedpolicy-implications that are not adequately supported by the data used for HCMestimation. More specifically, after finding that latent variables such asattitudes and perceptions are significantly related to choice behavior, authorsroutinely propose the development of policies that aim to influence theselatent variables and as a consequence choice behavior. Prime examples of such proposedpolicy options refer to information campaigns to increase the environmental consciousnessof travelers (and presumably as a consequence change their mobility behaviortowards choosing more sustainable travel options), or to improve travelers’perceived image of a particular travel mode like transit (and presumably as aconsequence change modal choice behavior). In this paper, we will argue that suchpolicy implications are not supported by the nature of the data in combinationwith the nature of the latent variables. Morespecifically, there are two reasons why there is a lack of theoretical supportfor the use of latent variables in HCMs as targets for transport policies: i)latent attitudes and perceptions are partly endogenous with respect to travelbehavior, precluding strong inference of causality; and ii) they are measuredat a single moment in time, precluding inference of within-person variation.Tostart with the first of these notions: there are at least three compellingreasons why latent variables and perceptions of the type used in HCMs are inmost cases to be treated as being partly endogenous with respect to the travelchoice itself. -         Firstly, both thelatent variable and the choice variable are likely to be jointly influenced bythe same (unmeasured) underlying factors, which causes endogeneity. -         Secondly, the travelchoice may influence the latent variable (as opposed to the other way around),due to learning effects.-         Thirdly, theempirically well-established theory of cognitive dissonance shows that peopleattempt to align their attitudes and perceptions with their actual choicebehavior. Again, such after-the-fact justifications imply that the causalrelation might run from the choice to the latent variable, implying endogeneityof the latter.Asecond issue which further complicates the derivation of policy implications targetinga latent variable, is that latent attitudes and perceptions are almost withoutexception measured (by means of indicators) at one point in time, as opposed toat several points in time. The reason why this cross-sectional nature of latentvariable measurements causes problems for the derivation of transport policyimplications, is fairly straightforward: when variables are observed in theform of cross-sectional data, only between-person comparisons based on differences in latent variables areallowed for, as opposed to within-person comparisons that are based on changes in the latent variable. Onlydifferences between individuals are observed, rather than changes for (or: variationwithin) the same individual. In other words, there is no observed covariation(of the latent variable and the choice-variable) at the individual level, andsince covariation is a prerequisite for causation, there can be no causalinference at the individual level.Thecombination of the partial endogeneity of the latent variable and thecross-sectional nature of the data implies that any policy derived from anestimated HCM which aims to change the travel behavior of individuals by changingtheir latent perceptions or attitudes, is built on quicksand.