International Choice Modelling Conference, International Choice Modelling Conference 2015

Font Size: 
Empirical Analysis of Hypothetical Bias in Stated Preference Experiments
Li Tang, Xia Luo, Yang Cheng, Fei Yang, Bin Ran

Last modified: 11 May 2015


As a major survey method for discrete choice analysis, Stated Preference (SP) experiments usually present sampled respondents with a number of different hypothetical situations, each consisting of a finite set of alternatives defined on a number of attribute dimensions. Respondents are required to make choices or state their preference among these situations. The deviation between ‘stated’ and ‘real’ choices or valuations is referred to as Hypothetical Bias (HB). This notion implies that, due to the hypothetical nature of SP experiments, people may behave inconsistently that they do not back up their stated preference with real commitments. Evidence of HB has been largely found and discussed in marketing and environmental area, such as economic valuation for endangered species, public space and air pollution. In transportation area, HB is usually identified in experiments estimating the Value of Time (VOT). Although efforts to study HB on marginal willingness to pay (WTP) on travel time have been made in recent years, only a very small number of literature can provide powerful results that the VOT in SP experiments are less than half amount of that in revealed preference (RP) data. Another motivation of this paper lies in the wide use of SP experiments for behavior analysis in transportation area. Especially for big cities in developing countries, new transportation modes and innovative traffic management methods, such as building a metro line or conducting congestion pricing, are now extensively introduced or discussed. Lack of former experience and practice, SP data seems to be the only source for predicting people’s reaction towards these changes. If HB really exists in such a big difference as reported in the literature, researchers may need to have a second thought in using SP data directly for model calibration and valuation.

This paper targets on providing an empirical study on HB about its existence and magnitude. We aim to present the following steps to the literature: first, considering most of the evidences on HB are explored in the aspect of economic analysis, we provide another dimension to represent it- in terms of model prediction results. A Selection-Calibration-Simulation method is proposed to convert individual choice data to aggregated market share of alternatives. Inputting a SP and a RP dataset respectively into the selected model structure, the chosen probability of each alternative can be obtained through the calibrated models and be compared in a uniform scenario using RP data as the benchmark. Second, a VOT model is presented to keep consistency with former successful studies in measuring HB. With our empirical travel mode choice data collected in China, we have a valuable chance to discuss whether HB varies with geographic distribution.

The travel mode choice data used for this analysis are collected on a corridor connecting suburb and urban areas in Chengdu, China. Four alternatives are involved in the study: car, taxi, bus and subway. Among them, the former three are long-existing transport modes and the subway, the No. 2 West Extended Line is officially operated on June 8th, 2013. The survey experiment is conducted in two stages. All the respondents are randomly sampled along the same corridor. First, a SP survey is taken from May 1st to June 4th, 2013. Respondents are required to state their travel mode choice in different choice situations. In this stage, the subway is still under construction. Thus it is regarded as a ‘hypothetical’ alternative in the choice set. The attribute levels are given based on previous study results as well as to preserve realistic estimates for the private and public transport alternatives. In the second stage, from April 9th to 18th, 2014, a RP survey was conducted to obtain the actual travel mode choices on the experiment corridor, along with the real value of attributes. Also, the social-demographic characteristic (SDC) data including gender, car ownership, age, income and trip purpose are collected in both surveys. Finally, 1440 observations are obtained from SP survey and 910 observations are obtained from RP survey. Seeing from the comparison diagrams of SDC variables, the SDC distribution of SP survey generally matches that of RP survey.


By comparing the model fit of Multinomial Logit (MNL) and Nested Logit, MNL model is selected as the basic model structure in this paper. After calibrating two separate MNL models based on SP and RP data, we ran the simulation using the attribute mean value from RP data as the input. The result shows a significant difference between the prediction results for each alternative when using models calibrated in SP/RP situation with the same input. In general, for the type of travel modes (i.e. public or private), people tend to behave consistently (75.524% vs. 75.399% in bus and subway, 24.475% vs. 24.601% in taxi and car). However, for a specific travel mode, an apparent increase of choices on bus can be observed while an obvious decrease of choice on subway are found. It indicates that in hypothetical situations, people may overestimate their preference on new type transit and neglect their reliance or habit on the traditional ones. Moreover, the calculation results of the VOT model further confirm the simulation result. For car users, the VOT from SP responses is about half that of RP, which is consistent with other research results mentioned earlier. Also, it reveals an interesting phenomenon that for bus, taxi and subway users, their SP VOT estimates are much lower than that of the RP estimates.

Conference registration is required in order to view papers.