International Choice Modelling Conference, International Choice Modelling Conference 2013

Font Size: 
Hybrid Choice Models for Decision Uncertainty: Implicitly or explicitly uncertain?
Thijs Dekker, Stephane Hess, Roy Brouwer, Marjan Hofkes

Last modified:  6 July 2013

Abstract


This paper presents a novel Bayesian Hybrid Choice Model dealing with decision uncertainty in stated choice studies. Treatment of self-reported choice certainty follow-up responses by choice modellers has been limited from a methodological perspective, relying on either a purely one-directional impact, making self-reported choice certainty a result of utility differences across the alternatives in the choice set without recognising that decision uncertainty itself may influence utility, or using self-reported choice certainty directly as an explanatory variable in the utility functions, with potential risk of endogeneity bias as well as measurement error.

Endogeneity arises when the alternatives in the choice task are close to each other in terms of their utility levels, then the choice task is perceived as complex and respondents report this in the follow-up question presented jointly with the choice task. Using the self-reported choice certainty response directly as an explanatory factor in the utility function therefore introduces correlation between the error term of the choice model and its explanatory variables. The presented hybrid choice model deals with potential endogeneity and measurement error by treating decision uncertainty as a latent variable which simultaneously affects the response to the choice task and the response to the self-reported choice certainty follow-up question.

 

The observed choices and follow-up responses are treated as independent variables forming respectively the choice model (mixed logit with varying scale parameter) and the measurement model (ordered probit) alleviating the endogeneity issue. Latent decision uncertainty forms the structural equation and is the connecting part between both models by introducing correlation. By tracing the simultaneous impact of latent decision uncertainty on the choice and measurement model, a more natural representation of the problem is presented compared to the sequential (or one-directional) views adopted before. In the model, the analyst learns about factors affecting decision certainty, while simultaneously controlling for its impact on decisions and welfare measures. A Bayesian model specification is adopted to reduce estimation time.

 

The empirical application of the model focuses on obtaining WTP estimates for flood risk reductions in the Netherlands. Flood risk safety in the Netherlands has so far been a full public good and private flood risk insurance schemes are not (yet) available. The ‘high impact – low probability’ characteristic of floods further adds to respondent unfamiliarity and lack of experience with floods and the presented trade-offs. Decision uncertainty is therefore likely to play a role in this stated choice survey.

 

Two hypothesis on the impact of decision uncertainty on the choice model are jointly tested in an empirical application on flood risk reductions in the Netherlands. First, certain respondents make less random decision consequently exhibiting a higher scale parameter. Second, uncertain respondents exhibit trade aversion making them more inclined to select the ‘opt-out’ option. Support is found for both hypotheses. Decision uncertainty decreases due to fatigue effects, but increases when being male or perceiving the presented choices as credible.

 


Conference registration is required in order to view papers.